SlideShare a Scribd company logo
MAML: Model-Agnostic Meta-Learning

for Fast Adaptation for Deep Networks
HyeongJoo Hwang

Jeonggwan Lee

Joonyoung Yi

2018. 11. 28.
Contents
1. Meta-Learning Problem Setup (Definition & Goal)

2. Prior Work

3. Model Agnostic Meta Learning

1. Characteristics

2. Intuition

4. Approach

1. Supervised Learning

2. Gradient of Gradient

3. Reinforcement Learning

5. Experiment

6. Discussion
Meta-Learning Problem Set-up
• Definition of meta-learning:

• Learner is trained by meta-learner to be able to learn on many different
tasks.

• Goal:

• Learner quickly learn new tasks from a small amount of new data.
1. Problem Set-up
Meta Supervised Learning
• Supervised learning:

• Meta supervised learning?
1. Problem Set-up
Meta Supervised Learning
1. Problem Set-up
Meta Supervised Learning
1. Problem Set-up
Meta Supervised Learning
1. Problem Set-up
Meta Supervised Learning
1. Problem Set-up
Meta Supervised Learning
1. Problem Set-up
Meta Supervised Learning
1. Problem Set-up
Meta Supervised Learning
1. Problem Set-up
Prior Work
• Learning an update function or update rule

• LSTM optimizer

(Learning to learn by gradient descent by gradient descent)

• Meta LSTM optimizer

(Optimization as a model for few-shot learning)

• Few shot (or meta) learning for specific tasks

• Generative modeling (Neural Statistician)

• Image classification (Matching Net., Prototypical Net.)

• Reinforcement learning 

(Benchmarking deep reinforcement learning for continuous control)
2. Prior Work
Prior Work
• Memory-augmented meta-learning

• Memory-augmented models on many tasks

(Meta-learning with memory-augmented neural networks)

• Parameter initialization of deep networks

• Exploring sensitivity while maintaining Internal representation

(Overcoming catastrophic forgetting in NN)

• Data-dependent Initializer

(Data-Dependent Initialization of CNN)

• learned initialization

(Gradient-based hyperparameter optimization through reversible learning)
2. Prior Work
Model-Agnostic Meta-Learning (MAML)
• Goal:

• Quickly adapt to new tasks on distribution 

with only small amount of data and 

with only a few gradient steps, 

even one gradient step.

• Learner:

• Learn a new task by using a single gradient step.

• Meta-learner:

• Learn a generalized parameter initialization of model.
3. Model-Agnostic Meta-Learning (MAML)
Characteristics of MAML
• The MAML learner’s weights are updated using the gradient, 

rather than a learned update.

• Not need additional parameters nor require a particular learner
architecture.

• Fast adaptability through good parameter initialization

• Explicitly optimizes to learn Internal representation

(i.e. suitable for many tasks).

• Maximizes sensitivity of new task losses to the model parameters.
3. Model-Agnostic Meta-Learning (MAML)
Characteristics of MAML
• Model-agnostic (No matter whatever model is)

• Classification & regression with differentiable losses, Policy gradient RL. 

• The model should be parameterized.

• No other assumptions on the form of the model.

• Task-agnostic (No matter whatever task is)

• Adopted all knowledge-transferable tasks.

• No other assumption is required.
3. Model-Agnostic Meta-Learning (MAML)
• Some internal representations

are more transferrable than others.

• Desired model parameter set is θ such that:

Applying one (or a small # of) gradient step to θ

on a new task will produce maximally effective behavior.



→ Find θ that commonly decreases loss of each task 

after adaptation.
Intuition of MAML
3. Model-Agnostic Meta-Learning (MAML)
Supervised Learning
• Notations

• Model: (Model-agnostic)

• Task distribution:

• For each Task: 

• Data distribution: q (K samples are drawn from q)

• Loss function:

• MSE for Regression

• Cross entropy for classification

• Any other differentiable loss functions can be used.
4. Approach
f✓<latexit sha1_base64="vYib4ronG48FFWnNXVHMYLu2Hrs=">AAAB7nicbVBNS8NAEJ3Ur1q/qh69LBbBU0lEUG8FL+KpgrGFNpTNdtMu3Wzi7kQooX/CiwcVr/4eb/4bt20O2vpg4PHeDDPzwlQKg6777ZRWVtfWN8qbla3tnd296v7Bg0kyzbjPEpnodkgNl0JxHwVK3k41p3EoeSscXU/91hPXRiTqHscpD2I6UCISjKKV2lGvi0OOtFetuXV3BrJMvILUoECzV/3q9hOWxVwhk9SYjuemGORUo2CSTyrdzPCUshEd8I6lisbcBPns3gk5sUqfRIm2pZDM1N8TOY2NGceh7YwpDs2iNxX/8zoZRpdBLlSaIVdsvijKJMGETJ8nfaE5Qzm2hDIt7K2EDammDG1EFRuCt/jyMvHP6ld17+681rgt0ijDERzDKXhwAQ24gSb4wEDCM7zCm/PovDjvzse8teQUM4fwB87nD4uEj90=</latexit><latexit sha1_base64="vYib4ronG48FFWnNXVHMYLu2Hrs=">AAAB7nicbVBNS8NAEJ3Ur1q/qh69LBbBU0lEUG8FL+KpgrGFNpTNdtMu3Wzi7kQooX/CiwcVr/4eb/4bt20O2vpg4PHeDDPzwlQKg6777ZRWVtfWN8qbla3tnd296v7Bg0kyzbjPEpnodkgNl0JxHwVK3k41p3EoeSscXU/91hPXRiTqHscpD2I6UCISjKKV2lGvi0OOtFetuXV3BrJMvILUoECzV/3q9hOWxVwhk9SYjuemGORUo2CSTyrdzPCUshEd8I6lisbcBPns3gk5sUqfRIm2pZDM1N8TOY2NGceh7YwpDs2iNxX/8zoZRpdBLlSaIVdsvijKJMGETJ8nfaE5Qzm2hDIt7K2EDammDG1EFRuCt/jyMvHP6ld17+681rgt0ijDERzDKXhwAQ24gSb4wEDCM7zCm/PovDjvzse8teQUM4fwB87nD4uEj90=</latexit><latexit sha1_base64="vYib4ronG48FFWnNXVHMYLu2Hrs=">AAAB7nicbVBNS8NAEJ3Ur1q/qh69LBbBU0lEUG8FL+KpgrGFNpTNdtMu3Wzi7kQooX/CiwcVr/4eb/4bt20O2vpg4PHeDDPzwlQKg6777ZRWVtfWN8qbla3tnd296v7Bg0kyzbjPEpnodkgNl0JxHwVK3k41p3EoeSscXU/91hPXRiTqHscpD2I6UCISjKKV2lGvi0OOtFetuXV3BrJMvILUoECzV/3q9hOWxVwhk9SYjuemGORUo2CSTyrdzPCUshEd8I6lisbcBPns3gk5sUqfRIm2pZDM1N8TOY2NGceh7YwpDs2iNxX/8zoZRpdBLlSaIVdsvijKJMGETJ8nfaE5Qzm2hDIt7K2EDammDG1EFRuCt/jyMvHP6ld17+681rgt0ijDERzDKXhwAQ24gSb4wEDCM7zCm/PovDjvzse8teQUM4fwB87nD4uEj90=</latexit><latexit sha1_base64="vYib4ronG48FFWnNXVHMYLu2Hrs=">AAAB7nicbVBNS8NAEJ3Ur1q/qh69LBbBU0lEUG8FL+KpgrGFNpTNdtMu3Wzi7kQooX/CiwcVr/4eb/4bt20O2vpg4PHeDDPzwlQKg6777ZRWVtfWN8qbla3tnd296v7Bg0kyzbjPEpnodkgNl0JxHwVK3k41p3EoeSscXU/91hPXRiTqHscpD2I6UCISjKKV2lGvi0OOtFetuXV3BrJMvILUoECzV/3q9hOWxVwhk9SYjuemGORUo2CSTyrdzPCUshEd8I6lisbcBPns3gk5sUqfRIm2pZDM1N8TOY2NGceh7YwpDs2iNxX/8zoZRpdBLlSaIVdsvijKJMGETJ8nfaE5Qzm2hDIt7K2EDammDG1EFRuCt/jyMvHP6ld17+681rgt0ijDERzDKXhwAQ24gSb4wEDCM7zCm/PovDjvzse8teQUM4fwB87nD4uEj90=</latexit>
T = {L(x1, y1, . . . , xK, yK), (x, y) ⇠ q}<latexit sha1_base64="3SPqX5Pd//SZfknsNJ7h81baRq4=">AAACH3icbVBNS0JBFJ1nX2ZfVss2lyRQEHkvAmsRCG0iWxhoBj55zBtHHZz30cy8UB7+lDb9lTYtKqKd/6Z56qK0AzMczrn3ztzjhpxJZZoTI7Wyura+kd7MbG3v7O5l9w/uZRAJQhsk4IF4cLGknPm0oZji9CEUFHsup013cJX4zScqJAv8uhqFtO3hns+6jGClJSdbrsMl2DHcQn7oWEUYJRfYnUDJIgydaqJUC0XtalYAWzIPHsEeO9mcWTKngGVizUkOzVFzst96KIk86ivCsZQtywxVO8ZCMcLpOGNHkoaYDHCPtjT1sUdlO54uOIYTrXSgGwh9fAVT9XdHjD0pR56rKz2s+nLRS8T/vFakuuftmPlhpKhPZg91Iw4qgCQt6DBBieIjTTARTP8VSB8LTJTONKNDsBZXXiaN09JFybo7y1Vu5mmk0RE6RnlkoTKqoGtUQw1E0DN6Re/ow3gx3oxP42tWmjLmPYfoD4zJD9DSnhc=</latexit><latexit sha1_base64="3SPqX5Pd//SZfknsNJ7h81baRq4=">AAACH3icbVBNS0JBFJ1nX2ZfVss2lyRQEHkvAmsRCG0iWxhoBj55zBtHHZz30cy8UB7+lDb9lTYtKqKd/6Z56qK0AzMczrn3ztzjhpxJZZoTI7Wyura+kd7MbG3v7O5l9w/uZRAJQhsk4IF4cLGknPm0oZji9CEUFHsup013cJX4zScqJAv8uhqFtO3hns+6jGClJSdbrsMl2DHcQn7oWEUYJRfYnUDJIgydaqJUC0XtalYAWzIPHsEeO9mcWTKngGVizUkOzVFzst96KIk86ivCsZQtywxVO8ZCMcLpOGNHkoaYDHCPtjT1sUdlO54uOIYTrXSgGwh9fAVT9XdHjD0pR56rKz2s+nLRS8T/vFakuuftmPlhpKhPZg91Iw4qgCQt6DBBieIjTTARTP8VSB8LTJTONKNDsBZXXiaN09JFybo7y1Vu5mmk0RE6RnlkoTKqoGtUQw1E0DN6Re/ow3gx3oxP42tWmjLmPYfoD4zJD9DSnhc=</latexit><latexit sha1_base64="3SPqX5Pd//SZfknsNJ7h81baRq4=">AAACH3icbVBNS0JBFJ1nX2ZfVss2lyRQEHkvAmsRCG0iWxhoBj55zBtHHZz30cy8UB7+lDb9lTYtKqKd/6Z56qK0AzMczrn3ztzjhpxJZZoTI7Wyura+kd7MbG3v7O5l9w/uZRAJQhsk4IF4cLGknPm0oZji9CEUFHsup013cJX4zScqJAv8uhqFtO3hns+6jGClJSdbrsMl2DHcQn7oWEUYJRfYnUDJIgydaqJUC0XtalYAWzIPHsEeO9mcWTKngGVizUkOzVFzst96KIk86ivCsZQtywxVO8ZCMcLpOGNHkoaYDHCPtjT1sUdlO54uOIYTrXSgGwh9fAVT9XdHjD0pR56rKz2s+nLRS8T/vFakuuftmPlhpKhPZg91Iw4qgCQt6DBBieIjTTARTP8VSB8LTJTONKNDsBZXXiaN09JFybo7y1Vu5mmk0RE6RnlkoTKqoGtUQw1E0DN6Re/ow3gx3oxP42tWmjLmPYfoD4zJD9DSnhc=</latexit><latexit sha1_base64="3SPqX5Pd//SZfknsNJ7h81baRq4=">AAACH3icbVBNS0JBFJ1nX2ZfVss2lyRQEHkvAmsRCG0iWxhoBj55zBtHHZz30cy8UB7+lDb9lTYtKqKd/6Z56qK0AzMczrn3ztzjhpxJZZoTI7Wyura+kd7MbG3v7O5l9w/uZRAJQhsk4IF4cLGknPm0oZji9CEUFHsup013cJX4zScqJAv8uhqFtO3hns+6jGClJSdbrsMl2DHcQn7oWEUYJRfYnUDJIgydaqJUC0XtalYAWzIPHsEeO9mcWTKngGVizUkOzVFzst96KIk86ivCsZQtywxVO8ZCMcLpOGNHkoaYDHCPtjT1sUdlO54uOIYTrXSgGwh9fAVT9XdHjD0pR56rKz2s+nLRS8T/vFakuuftmPlhpKhPZg91Iw4qgCQt6DBBieIjTTARTP8VSB8LTJTONKNDsBZXXiaN09JFybo7y1Vu5mmk0RE6RnlkoTKqoGtUQw1E0DN6Re/ow3gx3oxP42tWmjLmPYfoD4zJD9DSnhc=</latexit>
L<latexit sha1_base64="UIzmQdPm4e6dU3uEnO6VqVUUQ+8=">AAAB8XicbVBNSwMxFMzWr1q/qh69BIvgqeyKoN4KXkQ8VHBtYbuUbJptQ7PJkrwVytKf4cWDilf/jTf/jdl2D9o6EBhm3iPzJkoFN+C6305lZXVtfaO6Wdva3tndq+8fPBqVacp8qoTS3YgYJrhkPnAQrJtqRpJIsE40vi78zhPThiv5AJOUhQkZSh5zSsBKQS8hMKJE5HfTfr3hNt0Z8DLxStJAJdr9+ldvoGiWMAlUEGMCz00hzIkGTgWb1nqZYSmhYzJkgaWSJMyE+SzyFJ9YZYBjpe2TgGfq742cJMZMkshOFhHNoleI/3lBBvFlmHOZZsAknX8UZwKDwsX9eMA1oyAmlhCquc2K6YhoQsG2VLMleIsnLxP/rHnV9O7PG63bso0qOkLH6BR56AK10A1qIx9RpNAzekVvDjgvzrvzMR+tOOXOIfoD5/MH7biRPg==</latexit><latexit sha1_base64="UIzmQdPm4e6dU3uEnO6VqVUUQ+8=">AAAB8XicbVBNSwMxFMzWr1q/qh69BIvgqeyKoN4KXkQ8VHBtYbuUbJptQ7PJkrwVytKf4cWDilf/jTf/jdl2D9o6EBhm3iPzJkoFN+C6305lZXVtfaO6Wdva3tndq+8fPBqVacp8qoTS3YgYJrhkPnAQrJtqRpJIsE40vi78zhPThiv5AJOUhQkZSh5zSsBKQS8hMKJE5HfTfr3hNt0Z8DLxStJAJdr9+ldvoGiWMAlUEGMCz00hzIkGTgWb1nqZYSmhYzJkgaWSJMyE+SzyFJ9YZYBjpe2TgGfq742cJMZMkshOFhHNoleI/3lBBvFlmHOZZsAknX8UZwKDwsX9eMA1oyAmlhCquc2K6YhoQsG2VLMleIsnLxP/rHnV9O7PG63bso0qOkLH6BR56AK10A1qIx9RpNAzekVvDjgvzrvzMR+tOOXOIfoD5/MH7biRPg==</latexit><latexit sha1_base64="UIzmQdPm4e6dU3uEnO6VqVUUQ+8=">AAAB8XicbVBNSwMxFMzWr1q/qh69BIvgqeyKoN4KXkQ8VHBtYbuUbJptQ7PJkrwVytKf4cWDilf/jTf/jdl2D9o6EBhm3iPzJkoFN+C6305lZXVtfaO6Wdva3tndq+8fPBqVacp8qoTS3YgYJrhkPnAQrJtqRpJIsE40vi78zhPThiv5AJOUhQkZSh5zSsBKQS8hMKJE5HfTfr3hNt0Z8DLxStJAJdr9+ldvoGiWMAlUEGMCz00hzIkGTgWb1nqZYSmhYzJkgaWSJMyE+SzyFJ9YZYBjpe2TgGfq742cJMZMkshOFhHNoleI/3lBBvFlmHOZZsAknX8UZwKDwsX9eMA1oyAmlhCquc2K6YhoQsG2VLMleIsnLxP/rHnV9O7PG63bso0qOkLH6BR56AK10A1qIx9RpNAzekVvDjgvzrvzMR+tOOXOIfoD5/MH7biRPg==</latexit><latexit sha1_base64="UIzmQdPm4e6dU3uEnO6VqVUUQ+8=">AAAB8XicbVBNSwMxFMzWr1q/qh69BIvgqeyKoN4KXkQ8VHBtYbuUbJptQ7PJkrwVytKf4cWDilf/jTf/jdl2D9o6EBhm3iPzJkoFN+C6305lZXVtfaO6Wdva3tndq+8fPBqVacp8qoTS3YgYJrhkPnAQrJtqRpJIsE40vi78zhPThiv5AJOUhQkZSh5zSsBKQS8hMKJE5HfTfr3hNt0Z8DLxStJAJdr9+ldvoGiWMAlUEGMCz00hzIkGTgWb1nqZYSmhYzJkgaWSJMyE+SzyFJ9YZYBjpe2TgGfq742cJMZMkshOFhHNoleI/3lBBvFlmHOZZsAknX8UZwKDwsX9eMA1oyAmlhCquc2K6YhoQsG2VLMleIsnLxP/rHnV9O7PG63bso0qOkLH6BR56AK10A1qIx9RpNAzekVvDjgvzrvzMR+tOOXOIfoD5/MH7biRPg==</latexit>
p(T )<latexit sha1_base64="pt9rNGiWOedYqMAbHE8mujpf7lg=">AAAB9HicbVBNS8NAFHypX7V+VT16WSxCvZREBPVW8CKeKjS20May2W7apZtN2N0oJfR/ePGg4tUf481/4ybNQVsHFoaZ93iz48ecKW3b31ZpZXVtfaO8Wdna3tndq+4f3KsokYS6JOKR7PpYUc4EdTXTnHZjSXHoc9rxJ9eZ33mkUrFItPU0pl6IR4IFjGBtpIe43g+xHhPM0/bsdFCt2Q07B1omTkFqUKA1qH71hxFJQio04VipnmPH2kux1IxwOqv0E0VjTCZ4RHuGChxS5aV56hk6McoQBZE0T2iUq783UhwqNQ19M5llVIteJv7n9RIdXHopE3GiqSDzQ0HCkY5QVgEaMkmJ5lNDMJHMZEVkjCUm2hRVMSU4i19eJu5Z46rh3J3XmrdFG2U4gmOogwMX0IQbaIELBCQ8wyu8WU/Wi/VufcxHS1axcwh/YH3+AJi2kiU=</latexit><latexit sha1_base64="pt9rNGiWOedYqMAbHE8mujpf7lg=">AAAB9HicbVBNS8NAFHypX7V+VT16WSxCvZREBPVW8CKeKjS20May2W7apZtN2N0oJfR/ePGg4tUf481/4ybNQVsHFoaZ93iz48ecKW3b31ZpZXVtfaO8Wdna3tndq+4f3KsokYS6JOKR7PpYUc4EdTXTnHZjSXHoc9rxJ9eZ33mkUrFItPU0pl6IR4IFjGBtpIe43g+xHhPM0/bsdFCt2Q07B1omTkFqUKA1qH71hxFJQio04VipnmPH2kux1IxwOqv0E0VjTCZ4RHuGChxS5aV56hk6McoQBZE0T2iUq783UhwqNQ19M5llVIteJv7n9RIdXHopE3GiqSDzQ0HCkY5QVgEaMkmJ5lNDMJHMZEVkjCUm2hRVMSU4i19eJu5Z46rh3J3XmrdFG2U4gmOogwMX0IQbaIELBCQ8wyu8WU/Wi/VufcxHS1axcwh/YH3+AJi2kiU=</latexit><latexit sha1_base64="pt9rNGiWOedYqMAbHE8mujpf7lg=">AAAB9HicbVBNS8NAFHypX7V+VT16WSxCvZREBPVW8CKeKjS20May2W7apZtN2N0oJfR/ePGg4tUf481/4ybNQVsHFoaZ93iz48ecKW3b31ZpZXVtfaO8Wdna3tndq+4f3KsokYS6JOKR7PpYUc4EdTXTnHZjSXHoc9rxJ9eZ33mkUrFItPU0pl6IR4IFjGBtpIe43g+xHhPM0/bsdFCt2Q07B1omTkFqUKA1qH71hxFJQio04VipnmPH2kux1IxwOqv0E0VjTCZ4RHuGChxS5aV56hk6McoQBZE0T2iUq783UhwqNQ19M5llVIteJv7n9RIdXHopE3GiqSDzQ0HCkY5QVgEaMkmJ5lNDMJHMZEVkjCUm2hRVMSU4i19eJu5Z46rh3J3XmrdFG2U4gmOogwMX0IQbaIELBCQ8wyu8WU/Wi/VufcxHS1axcwh/YH3+AJi2kiU=</latexit><latexit sha1_base64="pt9rNGiWOedYqMAbHE8mujpf7lg=">AAAB9HicbVBNS8NAFHypX7V+VT16WSxCvZREBPVW8CKeKjS20May2W7apZtN2N0oJfR/ePGg4tUf481/4ybNQVsHFoaZ93iz48ecKW3b31ZpZXVtfaO8Wdna3tndq+4f3KsokYS6JOKR7PpYUc4EdTXTnHZjSXHoc9rxJ9eZ33mkUrFItPU0pl6IR4IFjGBtpIe43g+xHhPM0/bsdFCt2Q07B1omTkFqUKA1qH71hxFJQio04VipnmPH2kux1IxwOqv0E0VjTCZ4RHuGChxS5aV56hk6McoQBZE0T2iUq783UhwqNQ19M5llVIteJv7n9RIdXHopE3GiqSDzQ0HCkY5QVgEaMkmJ5lNDMJHMZEVkjCUm2hRVMSU4i19eJu5Z46rh3J3XmrdFG2U4gmOogwMX0IQbaIELBCQ8wyu8WU/Wi/VufcxHS1axcwh/YH3+AJi2kiU=</latexit>
Supervised Learning
4. Approach
• From line 10 in Algorithm 2, 



(Recall: )
Gradient of Gradient
4. Approach
• From line 10 in Algorithm 2, 



(Recall: )





( is differentiable)
Gradient of Gradient
4. Approach
L<latexit sha1_base64="UIzmQdPm4e6dU3uEnO6VqVUUQ+8=">AAAB8XicbVBNSwMxFMzWr1q/qh69BIvgqeyKoN4KXkQ8VHBtYbuUbJptQ7PJkrwVytKf4cWDilf/jTf/jdl2D9o6EBhm3iPzJkoFN+C6305lZXVtfaO6Wdva3tndq+8fPBqVacp8qoTS3YgYJrhkPnAQrJtqRpJIsE40vi78zhPThiv5AJOUhQkZSh5zSsBKQS8hMKJE5HfTfr3hNt0Z8DLxStJAJdr9+ldvoGiWMAlUEGMCz00hzIkGTgWb1nqZYSmhYzJkgaWSJMyE+SzyFJ9YZYBjpe2TgGfq742cJMZMkshOFhHNoleI/3lBBvFlmHOZZsAknX8UZwKDwsX9eMA1oyAmlhCquc2K6YhoQsG2VLMleIsnLxP/rHnV9O7PG63bso0qOkLH6BR56AK10A1qIx9RpNAzekVvDjgvzrvzMR+tOOXOIfoD5/MH7biRPg==</latexit><latexit sha1_base64="UIzmQdPm4e6dU3uEnO6VqVUUQ+8=">AAAB8XicbVBNSwMxFMzWr1q/qh69BIvgqeyKoN4KXkQ8VHBtYbuUbJptQ7PJkrwVytKf4cWDilf/jTf/jdl2D9o6EBhm3iPzJkoFN+C6305lZXVtfaO6Wdva3tndq+8fPBqVacp8qoTS3YgYJrhkPnAQrJtqRpJIsE40vi78zhPThiv5AJOUhQkZSh5zSsBKQS8hMKJE5HfTfr3hNt0Z8DLxStJAJdr9+ldvoGiWMAlUEGMCz00hzIkGTgWb1nqZYSmhYzJkgaWSJMyE+SzyFJ9YZYBjpe2TgGfq742cJMZMkshOFhHNoleI/3lBBvFlmHOZZsAknX8UZwKDwsX9eMA1oyAmlhCquc2K6YhoQsG2VLMleIsnLxP/rHnV9O7PG63bso0qOkLH6BR56AK10A1qIx9RpNAzekVvDjgvzrvzMR+tOOXOIfoD5/MH7biRPg==</latexit><latexit sha1_base64="UIzmQdPm4e6dU3uEnO6VqVUUQ+8=">AAAB8XicbVBNSwMxFMzWr1q/qh69BIvgqeyKoN4KXkQ8VHBtYbuUbJptQ7PJkrwVytKf4cWDilf/jTf/jdl2D9o6EBhm3iPzJkoFN+C6305lZXVtfaO6Wdva3tndq+8fPBqVacp8qoTS3YgYJrhkPnAQrJtqRpJIsE40vi78zhPThiv5AJOUhQkZSh5zSsBKQS8hMKJE5HfTfr3hNt0Z8DLxStJAJdr9+ldvoGiWMAlUEGMCz00hzIkGTgWb1nqZYSmhYzJkgaWSJMyE+SzyFJ9YZYBjpe2TgGfq742cJMZMkshOFhHNoleI/3lBBvFlmHOZZsAknX8UZwKDwsX9eMA1oyAmlhCquc2K6YhoQsG2VLMleIsnLxP/rHnV9O7PG63bso0qOkLH6BR56AK10A1qIx9RpNAzekVvDjgvzrvzMR+tOOXOIfoD5/MH7biRPg==</latexit><latexit sha1_base64="UIzmQdPm4e6dU3uEnO6VqVUUQ+8=">AAAB8XicbVBNSwMxFMzWr1q/qh69BIvgqeyKoN4KXkQ8VHBtYbuUbJptQ7PJkrwVytKf4cWDilf/jTf/jdl2D9o6EBhm3iPzJkoFN+C6305lZXVtfaO6Wdva3tndq+8fPBqVacp8qoTS3YgYJrhkPnAQrJtqRpJIsE40vi78zhPThiv5AJOUhQkZSh5zSsBKQS8hMKJE5HfTfr3hNt0Z8DLxStJAJdr9+ldvoGiWMAlUEGMCz00hzIkGTgWb1nqZYSmhYzJkgaWSJMyE+SzyFJ9YZYBjpe2TgGfq742cJMZMkshOFhHNoleI/3lBBvFlmHOZZsAknX8UZwKDwsX9eMA1oyAmlhCquc2K6YhoQsG2VLMleIsnLxP/rHnV9O7PG63bso0qOkLH6BR56AK10A1qIx9RpNAzekVvDjgvzrvzMR+tOOXOIfoD5/MH7biRPg==</latexit>
• From line 10 in Algorithm 2, 



(Recall: )





( is differentiable)
Gradient of Gradient
4. Approach
L<latexit sha1_base64="UIzmQdPm4e6dU3uEnO6VqVUUQ+8=">AAAB8XicbVBNSwMxFMzWr1q/qh69BIvgqeyKoN4KXkQ8VHBtYbuUbJptQ7PJkrwVytKf4cWDilf/jTf/jdl2D9o6EBhm3iPzJkoFN+C6305lZXVtfaO6Wdva3tndq+8fPBqVacp8qoTS3YgYJrhkPnAQrJtqRpJIsE40vi78zhPThiv5AJOUhQkZSh5zSsBKQS8hMKJE5HfTfr3hNt0Z8DLxStJAJdr9+ldvoGiWMAlUEGMCz00hzIkGTgWb1nqZYSmhYzJkgaWSJMyE+SzyFJ9YZYBjpe2TgGfq742cJMZMkshOFhHNoleI/3lBBvFlmHOZZsAknX8UZwKDwsX9eMA1oyAmlhCquc2K6YhoQsG2VLMleIsnLxP/rHnV9O7PG63bso0qOkLH6BR56AK10A1qIx9RpNAzekVvDjgvzrvzMR+tOOXOIfoD5/MH7biRPg==</latexit><latexit sha1_base64="UIzmQdPm4e6dU3uEnO6VqVUUQ+8=">AAAB8XicbVBNSwMxFMzWr1q/qh69BIvgqeyKoN4KXkQ8VHBtYbuUbJptQ7PJkrwVytKf4cWDilf/jTf/jdl2D9o6EBhm3iPzJkoFN+C6305lZXVtfaO6Wdva3tndq+8fPBqVacp8qoTS3YgYJrhkPnAQrJtqRpJIsE40vi78zhPThiv5AJOUhQkZSh5zSsBKQS8hMKJE5HfTfr3hNt0Z8DLxStJAJdr9+ldvoGiWMAlUEGMCz00hzIkGTgWb1nqZYSmhYzJkgaWSJMyE+SzyFJ9YZYBjpe2TgGfq742cJMZMkshOFhHNoleI/3lBBvFlmHOZZsAknX8UZwKDwsX9eMA1oyAmlhCquc2K6YhoQsG2VLMleIsnLxP/rHnV9O7PG63bso0qOkLH6BR56AK10A1qIx9RpNAzekVvDjgvzrvzMR+tOOXOIfoD5/MH7biRPg==</latexit><latexit sha1_base64="UIzmQdPm4e6dU3uEnO6VqVUUQ+8=">AAAB8XicbVBNSwMxFMzWr1q/qh69BIvgqeyKoN4KXkQ8VHBtYbuUbJptQ7PJkrwVytKf4cWDilf/jTf/jdl2D9o6EBhm3iPzJkoFN+C6305lZXVtfaO6Wdva3tndq+8fPBqVacp8qoTS3YgYJrhkPnAQrJtqRpJIsE40vi78zhPThiv5AJOUhQkZSh5zSsBKQS8hMKJE5HfTfr3hNt0Z8DLxStJAJdr9+ldvoGiWMAlUEGMCz00hzIkGTgWb1nqZYSmhYzJkgaWSJMyE+SzyFJ9YZYBjpe2TgGfq742cJMZMkshOFhHNoleI/3lBBvFlmHOZZsAknX8UZwKDwsX9eMA1oyAmlhCquc2K6YhoQsG2VLMleIsnLxP/rHnV9O7PG63bso0qOkLH6BR56AK10A1qIx9RpNAzekVvDjgvzrvzMR+tOOXOIfoD5/MH7biRPg==</latexit><latexit sha1_base64="UIzmQdPm4e6dU3uEnO6VqVUUQ+8=">AAAB8XicbVBNSwMxFMzWr1q/qh69BIvgqeyKoN4KXkQ8VHBtYbuUbJptQ7PJkrwVytKf4cWDilf/jTf/jdl2D9o6EBhm3iPzJkoFN+C6305lZXVtfaO6Wdva3tndq+8fPBqVacp8qoTS3YgYJrhkPnAQrJtqRpJIsE40vi78zhPThiv5AJOUhQkZSh5zSsBKQS8hMKJE5HfTfr3hNt0Z8DLxStJAJdr9+ldvoGiWMAlUEGMCz00hzIkGTgWb1nqZYSmhYzJkgaWSJMyE+SzyFJ9YZYBjpe2TgGfq742cJMZMkshOFhHNoleI/3lBBvFlmHOZZsAknX8UZwKDwsX9eMA1oyAmlhCquc2K6YhoQsG2VLMleIsnLxP/rHnV9O7PG63bso0qOkLH6BR56AK10A1qIx9RpNAzekVvDjgvzrvzMR+tOOXOIfoD5/MH7biRPg==</latexit>
• From line 10 in Algorithm 2, 



(Recall: )





( is differentiable)
Gradient of Gradient
4. Approach
L<latexit sha1_base64="UIzmQdPm4e6dU3uEnO6VqVUUQ+8=">AAAB8XicbVBNSwMxFMzWr1q/qh69BIvgqeyKoN4KXkQ8VHBtYbuUbJptQ7PJkrwVytKf4cWDilf/jTf/jdl2D9o6EBhm3iPzJkoFN+C6305lZXVtfaO6Wdva3tndq+8fPBqVacp8qoTS3YgYJrhkPnAQrJtqRpJIsE40vi78zhPThiv5AJOUhQkZSh5zSsBKQS8hMKJE5HfTfr3hNt0Z8DLxStJAJdr9+ldvoGiWMAlUEGMCz00hzIkGTgWb1nqZYSmhYzJkgaWSJMyE+SzyFJ9YZYBjpe2TgGfq742cJMZMkshOFhHNoleI/3lBBvFlmHOZZsAknX8UZwKDwsX9eMA1oyAmlhCquc2K6YhoQsG2VLMleIsnLxP/rHnV9O7PG63bso0qOkLH6BR56AK10A1qIx9RpNAzekVvDjgvzrvzMR+tOOXOIfoD5/MH7biRPg==</latexit><latexit sha1_base64="UIzmQdPm4e6dU3uEnO6VqVUUQ+8=">AAAB8XicbVBNSwMxFMzWr1q/qh69BIvgqeyKoN4KXkQ8VHBtYbuUbJptQ7PJkrwVytKf4cWDilf/jTf/jdl2D9o6EBhm3iPzJkoFN+C6305lZXVtfaO6Wdva3tndq+8fPBqVacp8qoTS3YgYJrhkPnAQrJtqRpJIsE40vi78zhPThiv5AJOUhQkZSh5zSsBKQS8hMKJE5HfTfr3hNt0Z8DLxStJAJdr9+ldvoGiWMAlUEGMCz00hzIkGTgWb1nqZYSmhYzJkgaWSJMyE+SzyFJ9YZYBjpe2TgGfq742cJMZMkshOFhHNoleI/3lBBvFlmHOZZsAknX8UZwKDwsX9eMA1oyAmlhCquc2K6YhoQsG2VLMleIsnLxP/rHnV9O7PG63bso0qOkLH6BR56AK10A1qIx9RpNAzekVvDjgvzrvzMR+tOOXOIfoD5/MH7biRPg==</latexit><latexit sha1_base64="UIzmQdPm4e6dU3uEnO6VqVUUQ+8=">AAAB8XicbVBNSwMxFMzWr1q/qh69BIvgqeyKoN4KXkQ8VHBtYbuUbJptQ7PJkrwVytKf4cWDilf/jTf/jdl2D9o6EBhm3iPzJkoFN+C6305lZXVtfaO6Wdva3tndq+8fPBqVacp8qoTS3YgYJrhkPnAQrJtqRpJIsE40vi78zhPThiv5AJOUhQkZSh5zSsBKQS8hMKJE5HfTfr3hNt0Z8DLxStJAJdr9+ldvoGiWMAlUEGMCz00hzIkGTgWb1nqZYSmhYzJkgaWSJMyE+SzyFJ9YZYBjpe2TgGfq742cJMZMkshOFhHNoleI/3lBBvFlmHOZZsAknX8UZwKDwsX9eMA1oyAmlhCquc2K6YhoQsG2VLMleIsnLxP/rHnV9O7PG63bso0qOkLH6BR56AK10A1qIx9RpNAzekVvDjgvzrvzMR+tOOXOIfoD5/MH7biRPg==</latexit><latexit sha1_base64="UIzmQdPm4e6dU3uEnO6VqVUUQ+8=">AAAB8XicbVBNSwMxFMzWr1q/qh69BIvgqeyKoN4KXkQ8VHBtYbuUbJptQ7PJkrwVytKf4cWDilf/jTf/jdl2D9o6EBhm3iPzJkoFN+C6305lZXVtfaO6Wdva3tndq+8fPBqVacp8qoTS3YgYJrhkPnAQrJtqRpJIsE40vi78zhPThiv5AJOUhQkZSh5zSsBKQS8hMKJE5HfTfr3hNt0Z8DLxStJAJdr9+ldvoGiWMAlUEGMCz00hzIkGTgWb1nqZYSmhYzJkgaWSJMyE+SzyFJ9YZYBjpe2TgGfq742cJMZMkshOFhHNoleI/3lBBvFlmHOZZsAknX8UZwKDwsX9eMA1oyAmlhCquc2K6YhoQsG2VLMleIsnLxP/rHnV9O7PG63bso0qOkLH6BR56AK10A1qIx9RpNAzekVvDjgvzrvzMR+tOOXOIfoD5/MH7biRPg==</latexit>
• From line 10 in Algorithm 2, 



(Recall: )





( is differentiable)
Gradient of Gradient
4. Approach
Calculation of Hessian matrix is required.

L<latexit sha1_base64="UIzmQdPm4e6dU3uEnO6VqVUUQ+8=">AAAB8XicbVBNSwMxFMzWr1q/qh69BIvgqeyKoN4KXkQ8VHBtYbuUbJptQ7PJkrwVytKf4cWDilf/jTf/jdl2D9o6EBhm3iPzJkoFN+C6305lZXVtfaO6Wdva3tndq+8fPBqVacp8qoTS3YgYJrhkPnAQrJtqRpJIsE40vi78zhPThiv5AJOUhQkZSh5zSsBKQS8hMKJE5HfTfr3hNt0Z8DLxStJAJdr9+ldvoGiWMAlUEGMCz00hzIkGTgWb1nqZYSmhYzJkgaWSJMyE+SzyFJ9YZYBjpe2TgGfq742cJMZMkshOFhHNoleI/3lBBvFlmHOZZsAknX8UZwKDwsX9eMA1oyAmlhCquc2K6YhoQsG2VLMleIsnLxP/rHnV9O7PG63bso0qOkLH6BR56AK10A1qIx9RpNAzekVvDjgvzrvzMR+tOOXOIfoD5/MH7biRPg==</latexit><latexit sha1_base64="UIzmQdPm4e6dU3uEnO6VqVUUQ+8=">AAAB8XicbVBNSwMxFMzWr1q/qh69BIvgqeyKoN4KXkQ8VHBtYbuUbJptQ7PJkrwVytKf4cWDilf/jTf/jdl2D9o6EBhm3iPzJkoFN+C6305lZXVtfaO6Wdva3tndq+8fPBqVacp8qoTS3YgYJrhkPnAQrJtqRpJIsE40vi78zhPThiv5AJOUhQkZSh5zSsBKQS8hMKJE5HfTfr3hNt0Z8DLxStJAJdr9+ldvoGiWMAlUEGMCz00hzIkGTgWb1nqZYSmhYzJkgaWSJMyE+SzyFJ9YZYBjpe2TgGfq742cJMZMkshOFhHNoleI/3lBBvFlmHOZZsAknX8UZwKDwsX9eMA1oyAmlhCquc2K6YhoQsG2VLMleIsnLxP/rHnV9O7PG63bso0qOkLH6BR56AK10A1qIx9RpNAzekVvDjgvzrvzMR+tOOXOIfoD5/MH7biRPg==</latexit><latexit sha1_base64="UIzmQdPm4e6dU3uEnO6VqVUUQ+8=">AAAB8XicbVBNSwMxFMzWr1q/qh69BIvgqeyKoN4KXkQ8VHBtYbuUbJptQ7PJkrwVytKf4cWDilf/jTf/jdl2D9o6EBhm3iPzJkoFN+C6305lZXVtfaO6Wdva3tndq+8fPBqVacp8qoTS3YgYJrhkPnAQrJtqRpJIsE40vi78zhPThiv5AJOUhQkZSh5zSsBKQS8hMKJE5HfTfr3hNt0Z8DLxStJAJdr9+ldvoGiWMAlUEGMCz00hzIkGTgWb1nqZYSmhYzJkgaWSJMyE+SzyFJ9YZYBjpe2TgGfq742cJMZMkshOFhHNoleI/3lBBvFlmHOZZsAknX8UZwKDwsX9eMA1oyAmlhCquc2K6YhoQsG2VLMleIsnLxP/rHnV9O7PG63bso0qOkLH6BR56AK10A1qIx9RpNAzekVvDjgvzrvzMR+tOOXOIfoD5/MH7biRPg==</latexit><latexit sha1_base64="UIzmQdPm4e6dU3uEnO6VqVUUQ+8=">AAAB8XicbVBNSwMxFMzWr1q/qh69BIvgqeyKoN4KXkQ8VHBtYbuUbJptQ7PJkrwVytKf4cWDilf/jTf/jdl2D9o6EBhm3iPzJkoFN+C6305lZXVtfaO6Wdva3tndq+8fPBqVacp8qoTS3YgYJrhkPnAQrJtqRpJIsE40vi78zhPThiv5AJOUhQkZSh5zSsBKQS8hMKJE5HfTfr3hNt0Z8DLxStJAJdr9+ldvoGiWMAlUEGMCz00hzIkGTgWb1nqZYSmhYzJkgaWSJMyE+SzyFJ9YZYBjpe2TgGfq742cJMZMkshOFhHNoleI/3lBBvFlmHOZZsAknX8UZwKDwsX9eMA1oyAmlhCquc2K6YhoQsG2VLMleIsnLxP/rHnV9O7PG63bso0qOkLH6BR56AK10A1qIx9RpNAzekVvDjgvzrvzMR+tOOXOIfoD5/MH7biRPg==</latexit>
• From line 10 in Algorithm 2, 



(Recall: )





( is differentiable)
Gradient of Gradient
4. Approach
Calculation of Hessian matrix is required.

→ MAML suggest 1st order approximation.
L<latexit sha1_base64="UIzmQdPm4e6dU3uEnO6VqVUUQ+8=">AAAB8XicbVBNSwMxFMzWr1q/qh69BIvgqeyKoN4KXkQ8VHBtYbuUbJptQ7PJkrwVytKf4cWDilf/jTf/jdl2D9o6EBhm3iPzJkoFN+C6305lZXVtfaO6Wdva3tndq+8fPBqVacp8qoTS3YgYJrhkPnAQrJtqRpJIsE40vi78zhPThiv5AJOUhQkZSh5zSsBKQS8hMKJE5HfTfr3hNt0Z8DLxStJAJdr9+ldvoGiWMAlUEGMCz00hzIkGTgWb1nqZYSmhYzJkgaWSJMyE+SzyFJ9YZYBjpe2TgGfq742cJMZMkshOFhHNoleI/3lBBvFlmHOZZsAknX8UZwKDwsX9eMA1oyAmlhCquc2K6YhoQsG2VLMleIsnLxP/rHnV9O7PG63bso0qOkLH6BR56AK10A1qIx9RpNAzekVvDjgvzrvzMR+tOOXOIfoD5/MH7biRPg==</latexit><latexit sha1_base64="UIzmQdPm4e6dU3uEnO6VqVUUQ+8=">AAAB8XicbVBNSwMxFMzWr1q/qh69BIvgqeyKoN4KXkQ8VHBtYbuUbJptQ7PJkrwVytKf4cWDilf/jTf/jdl2D9o6EBhm3iPzJkoFN+C6305lZXVtfaO6Wdva3tndq+8fPBqVacp8qoTS3YgYJrhkPnAQrJtqRpJIsE40vi78zhPThiv5AJOUhQkZSh5zSsBKQS8hMKJE5HfTfr3hNt0Z8DLxStJAJdr9+ldvoGiWMAlUEGMCz00hzIkGTgWb1nqZYSmhYzJkgaWSJMyE+SzyFJ9YZYBjpe2TgGfq742cJMZMkshOFhHNoleI/3lBBvFlmHOZZsAknX8UZwKDwsX9eMA1oyAmlhCquc2K6YhoQsG2VLMleIsnLxP/rHnV9O7PG63bso0qOkLH6BR56AK10A1qIx9RpNAzekVvDjgvzrvzMR+tOOXOIfoD5/MH7biRPg==</latexit><latexit sha1_base64="UIzmQdPm4e6dU3uEnO6VqVUUQ+8=">AAAB8XicbVBNSwMxFMzWr1q/qh69BIvgqeyKoN4KXkQ8VHBtYbuUbJptQ7PJkrwVytKf4cWDilf/jTf/jdl2D9o6EBhm3iPzJkoFN+C6305lZXVtfaO6Wdva3tndq+8fPBqVacp8qoTS3YgYJrhkPnAQrJtqRpJIsE40vi78zhPThiv5AJOUhQkZSh5zSsBKQS8hMKJE5HfTfr3hNt0Z8DLxStJAJdr9+ldvoGiWMAlUEGMCz00hzIkGTgWb1nqZYSmhYzJkgaWSJMyE+SzyFJ9YZYBjpe2TgGfq742cJMZMkshOFhHNoleI/3lBBvFlmHOZZsAknX8UZwKDwsX9eMA1oyAmlhCquc2K6YhoQsG2VLMleIsnLxP/rHnV9O7PG63bso0qOkLH6BR56AK10A1qIx9RpNAzekVvDjgvzrvzMR+tOOXOIfoD5/MH7biRPg==</latexit><latexit sha1_base64="UIzmQdPm4e6dU3uEnO6VqVUUQ+8=">AAAB8XicbVBNSwMxFMzWr1q/qh69BIvgqeyKoN4KXkQ8VHBtYbuUbJptQ7PJkrwVytKf4cWDilf/jTf/jdl2D9o6EBhm3iPzJkoFN+C6305lZXVtfaO6Wdva3tndq+8fPBqVacp8qoTS3YgYJrhkPnAQrJtqRpJIsE40vi78zhPThiv5AJOUhQkZSh5zSsBKQS8hMKJE5HfTfr3hNt0Z8DLxStJAJdr9+ldvoGiWMAlUEGMCz00hzIkGTgWb1nqZYSmhYzJkgaWSJMyE+SzyFJ9YZYBjpe2TgGfq742cJMZMkshOFhHNoleI/3lBBvFlmHOZZsAknX8UZwKDwsX9eMA1oyAmlhCquc2K6YhoQsG2VLMleIsnLxP/rHnV9O7PG63bso0qOkLH6BR56AK10A1qIx9RpNAzekVvDjgvzrvzMR+tOOXOIfoD5/MH7biRPg==</latexit>
1st Order Approximation
• Update rule of MAML:
4. Approach
1st Order Approximation
• Update rule of MAML:
4. Approach
This part needs calculating Hessian.

Hessian makes MAML be slow.
1st Order Approximation
• Update rule of MAML:

• Update rule of MAML with 1st order approximation:
4. Approach
1st Order Approximation
• Update rule of MAML:

• Update rule of MAML with 1st order approximation:
4. Approach
(Regard as constant)
• From line 10 in Algorithm 2, 



(Recall: )





( is differentiable)
Recall: Gradient of Gradient
4. Approach
In 1st order approximation, 

we regard this as identity matrix I.
L<latexit sha1_base64="UIzmQdPm4e6dU3uEnO6VqVUUQ+8=">AAAB8XicbVBNSwMxFMzWr1q/qh69BIvgqeyKoN4KXkQ8VHBtYbuUbJptQ7PJkrwVytKf4cWDilf/jTf/jdl2D9o6EBhm3iPzJkoFN+C6305lZXVtfaO6Wdva3tndq+8fPBqVacp8qoTS3YgYJrhkPnAQrJtqRpJIsE40vi78zhPThiv5AJOUhQkZSh5zSsBKQS8hMKJE5HfTfr3hNt0Z8DLxStJAJdr9+ldvoGiWMAlUEGMCz00hzIkGTgWb1nqZYSmhYzJkgaWSJMyE+SzyFJ9YZYBjpe2TgGfq742cJMZMkshOFhHNoleI/3lBBvFlmHOZZsAknX8UZwKDwsX9eMA1oyAmlhCquc2K6YhoQsG2VLMleIsnLxP/rHnV9O7PG63bso0qOkLH6BR56AK10A1qIx9RpNAzekVvDjgvzrvzMR+tOOXOIfoD5/MH7biRPg==</latexit><latexit sha1_base64="UIzmQdPm4e6dU3uEnO6VqVUUQ+8=">AAAB8XicbVBNSwMxFMzWr1q/qh69BIvgqeyKoN4KXkQ8VHBtYbuUbJptQ7PJkrwVytKf4cWDilf/jTf/jdl2D9o6EBhm3iPzJkoFN+C6305lZXVtfaO6Wdva3tndq+8fPBqVacp8qoTS3YgYJrhkPnAQrJtqRpJIsE40vi78zhPThiv5AJOUhQkZSh5zSsBKQS8hMKJE5HfTfr3hNt0Z8DLxStJAJdr9+ldvoGiWMAlUEGMCz00hzIkGTgWb1nqZYSmhYzJkgaWSJMyE+SzyFJ9YZYBjpe2TgGfq742cJMZMkshOFhHNoleI/3lBBvFlmHOZZsAknX8UZwKDwsX9eMA1oyAmlhCquc2K6YhoQsG2VLMleIsnLxP/rHnV9O7PG63bso0qOkLH6BR56AK10A1qIx9RpNAzekVvDjgvzrvzMR+tOOXOIfoD5/MH7biRPg==</latexit><latexit sha1_base64="UIzmQdPm4e6dU3uEnO6VqVUUQ+8=">AAAB8XicbVBNSwMxFMzWr1q/qh69BIvgqeyKoN4KXkQ8VHBtYbuUbJptQ7PJkrwVytKf4cWDilf/jTf/jdl2D9o6EBhm3iPzJkoFN+C6305lZXVtfaO6Wdva3tndq+8fPBqVacp8qoTS3YgYJrhkPnAQrJtqRpJIsE40vi78zhPThiv5AJOUhQkZSh5zSsBKQS8hMKJE5HfTfr3hNt0Z8DLxStJAJdr9+ldvoGiWMAlUEGMCz00hzIkGTgWb1nqZYSmhYzJkgaWSJMyE+SzyFJ9YZYBjpe2TgGfq742cJMZMkshOFhHNoleI/3lBBvFlmHOZZsAknX8UZwKDwsX9eMA1oyAmlhCquc2K6YhoQsG2VLMleIsnLxP/rHnV9O7PG63bso0qOkLH6BR56AK10A1qIx9RpNAzekVvDjgvzrvzMR+tOOXOIfoD5/MH7biRPg==</latexit><latexit sha1_base64="UIzmQdPm4e6dU3uEnO6VqVUUQ+8=">AAAB8XicbVBNSwMxFMzWr1q/qh69BIvgqeyKoN4KXkQ8VHBtYbuUbJptQ7PJkrwVytKf4cWDilf/jTf/jdl2D9o6EBhm3iPzJkoFN+C6305lZXVtfaO6Wdva3tndq+8fPBqVacp8qoTS3YgYJrhkPnAQrJtqRpJIsE40vi78zhPThiv5AJOUhQkZSh5zSsBKQS8hMKJE5HfTfr3hNt0Z8DLxStJAJdr9+ldvoGiWMAlUEGMCz00hzIkGTgWb1nqZYSmhYzJkgaWSJMyE+SzyFJ9YZYBjpe2TgGfq742cJMZMkshOFhHNoleI/3lBBvFlmHOZZsAknX8UZwKDwsX9eMA1oyAmlhCquc2K6YhoQsG2VLMleIsnLxP/rHnV9O7PG63bso0qOkLH6BR56AK10A1qIx9RpNAzekVvDjgvzrvzMR+tOOXOIfoD5/MH7biRPg==</latexit>
Reinforcement Learning
• Notations

• Policy of an agent: such that a_t sim f_theta (x_t)

• Task distribution:

• For each Task: 

• Episode: 

• Transition distribution: q (K trajectories are drawn from q and f)

• Episode length: H

• Loss function: = expectation of sum of rewards
4. Approach
f✓<latexit sha1_base64="vYib4ronG48FFWnNXVHMYLu2Hrs=">AAAB7nicbVBNS8NAEJ3Ur1q/qh69LBbBU0lEUG8FL+KpgrGFNpTNdtMu3Wzi7kQooX/CiwcVr/4eb/4bt20O2vpg4PHeDDPzwlQKg6777ZRWVtfWN8qbla3tnd296v7Bg0kyzbjPEpnodkgNl0JxHwVK3k41p3EoeSscXU/91hPXRiTqHscpD2I6UCISjKKV2lGvi0OOtFetuXV3BrJMvILUoECzV/3q9hOWxVwhk9SYjuemGORUo2CSTyrdzPCUshEd8I6lisbcBPns3gk5sUqfRIm2pZDM1N8TOY2NGceh7YwpDs2iNxX/8zoZRpdBLlSaIVdsvijKJMGETJ8nfaE5Qzm2hDIt7K2EDammDG1EFRuCt/jyMvHP6ld17+681rgt0ijDERzDKXhwAQ24gSb4wEDCM7zCm/PovDjvzse8teQUM4fwB87nD4uEj90=</latexit><latexit sha1_base64="vYib4ronG48FFWnNXVHMYLu2Hrs=">AAAB7nicbVBNS8NAEJ3Ur1q/qh69LBbBU0lEUG8FL+KpgrGFNpTNdtMu3Wzi7kQooX/CiwcVr/4eb/4bt20O2vpg4PHeDDPzwlQKg6777ZRWVtfWN8qbla3tnd296v7Bg0kyzbjPEpnodkgNl0JxHwVK3k41p3EoeSscXU/91hPXRiTqHscpD2I6UCISjKKV2lGvi0OOtFetuXV3BrJMvILUoECzV/3q9hOWxVwhk9SYjuemGORUo2CSTyrdzPCUshEd8I6lisbcBPns3gk5sUqfRIm2pZDM1N8TOY2NGceh7YwpDs2iNxX/8zoZRpdBLlSaIVdsvijKJMGETJ8nfaE5Qzm2hDIt7K2EDammDG1EFRuCt/jyMvHP6ld17+681rgt0ijDERzDKXhwAQ24gSb4wEDCM7zCm/PovDjvzse8teQUM4fwB87nD4uEj90=</latexit><latexit sha1_base64="vYib4ronG48FFWnNXVHMYLu2Hrs=">AAAB7nicbVBNS8NAEJ3Ur1q/qh69LBbBU0lEUG8FL+KpgrGFNpTNdtMu3Wzi7kQooX/CiwcVr/4eb/4bt20O2vpg4PHeDDPzwlQKg6777ZRWVtfWN8qbla3tnd296v7Bg0kyzbjPEpnodkgNl0JxHwVK3k41p3EoeSscXU/91hPXRiTqHscpD2I6UCISjKKV2lGvi0OOtFetuXV3BrJMvILUoECzV/3q9hOWxVwhk9SYjuemGORUo2CSTyrdzPCUshEd8I6lisbcBPns3gk5sUqfRIm2pZDM1N8TOY2NGceh7YwpDs2iNxX/8zoZRpdBLlSaIVdsvijKJMGETJ8nfaE5Qzm2hDIt7K2EDammDG1EFRuCt/jyMvHP6ld17+681rgt0ijDERzDKXhwAQ24gSb4wEDCM7zCm/PovDjvzse8teQUM4fwB87nD4uEj90=</latexit><latexit sha1_base64="vYib4ronG48FFWnNXVHMYLu2Hrs=">AAAB7nicbVBNS8NAEJ3Ur1q/qh69LBbBU0lEUG8FL+KpgrGFNpTNdtMu3Wzi7kQooX/CiwcVr/4eb/4bt20O2vpg4PHeDDPzwlQKg6777ZRWVtfWN8qbla3tnd296v7Bg0kyzbjPEpnodkgNl0JxHwVK3k41p3EoeSscXU/91hPXRiTqHscpD2I6UCISjKKV2lGvi0OOtFetuXV3BrJMvILUoECzV/3q9hOWxVwhk9SYjuemGORUo2CSTyrdzPCUshEd8I6lisbcBPns3gk5sUqfRIm2pZDM1N8TOY2NGceh7YwpDs2iNxX/8zoZRpdBLlSaIVdsvijKJMGETJ8nfaE5Qzm2hDIt7K2EDammDG1EFRuCt/jyMvHP6ld17+681rgt0ijDERzDKXhwAQ24gSb4wEDCM7zCm/PovDjvzse8teQUM4fwB87nD4uEj90=</latexit>
p(T )<latexit sha1_base64="pt9rNGiWOedYqMAbHE8mujpf7lg=">AAAB9HicbVBNS8NAFHypX7V+VT16WSxCvZREBPVW8CKeKjS20May2W7apZtN2N0oJfR/ePGg4tUf481/4ybNQVsHFoaZ93iz48ecKW3b31ZpZXVtfaO8Wdna3tndq+4f3KsokYS6JOKR7PpYUc4EdTXTnHZjSXHoc9rxJ9eZ33mkUrFItPU0pl6IR4IFjGBtpIe43g+xHhPM0/bsdFCt2Q07B1omTkFqUKA1qH71hxFJQio04VipnmPH2kux1IxwOqv0E0VjTCZ4RHuGChxS5aV56hk6McoQBZE0T2iUq783UhwqNQ19M5llVIteJv7n9RIdXHopE3GiqSDzQ0HCkY5QVgEaMkmJ5lNDMJHMZEVkjCUm2hRVMSU4i19eJu5Z46rh3J3XmrdFG2U4gmOogwMX0IQbaIELBCQ8wyu8WU/Wi/VufcxHS1axcwh/YH3+AJi2kiU=</latexit><latexit sha1_base64="pt9rNGiWOedYqMAbHE8mujpf7lg=">AAAB9HicbVBNS8NAFHypX7V+VT16WSxCvZREBPVW8CKeKjS20May2W7apZtN2N0oJfR/ePGg4tUf481/4ybNQVsHFoaZ93iz48ecKW3b31ZpZXVtfaO8Wdna3tndq+4f3KsokYS6JOKR7PpYUc4EdTXTnHZjSXHoc9rxJ9eZ33mkUrFItPU0pl6IR4IFjGBtpIe43g+xHhPM0/bsdFCt2Q07B1omTkFqUKA1qH71hxFJQio04VipnmPH2kux1IxwOqv0E0VjTCZ4RHuGChxS5aV56hk6McoQBZE0T2iUq783UhwqNQ19M5llVIteJv7n9RIdXHopE3GiqSDzQ0HCkY5QVgEaMkmJ5lNDMJHMZEVkjCUm2hRVMSU4i19eJu5Z46rh3J3XmrdFG2U4gmOogwMX0IQbaIELBCQ8wyu8WU/Wi/VufcxHS1axcwh/YH3+AJi2kiU=</latexit><latexit sha1_base64="pt9rNGiWOedYqMAbHE8mujpf7lg=">AAAB9HicbVBNS8NAFHypX7V+VT16WSxCvZREBPVW8CKeKjS20May2W7apZtN2N0oJfR/ePGg4tUf481/4ybNQVsHFoaZ93iz48ecKW3b31ZpZXVtfaO8Wdna3tndq+4f3KsokYS6JOKR7PpYUc4EdTXTnHZjSXHoc9rxJ9eZ33mkUrFItPU0pl6IR4IFjGBtpIe43g+xHhPM0/bsdFCt2Q07B1omTkFqUKA1qH71hxFJQio04VipnmPH2kux1IxwOqv0E0VjTCZ4RHuGChxS5aV56hk6McoQBZE0T2iUq783UhwqNQ19M5llVIteJv7n9RIdXHopE3GiqSDzQ0HCkY5QVgEaMkmJ5lNDMJHMZEVkjCUm2hRVMSU4i19eJu5Z46rh3J3XmrdFG2U4gmOogwMX0IQbaIELBCQ8wyu8WU/Wi/VufcxHS1axcwh/YH3+AJi2kiU=</latexit><latexit sha1_base64="pt9rNGiWOedYqMAbHE8mujpf7lg=">AAAB9HicbVBNS8NAFHypX7V+VT16WSxCvZREBPVW8CKeKjS20May2W7apZtN2N0oJfR/ePGg4tUf481/4ybNQVsHFoaZ93iz48ecKW3b31ZpZXVtfaO8Wdna3tndq+4f3KsokYS6JOKR7PpYUc4EdTXTnHZjSXHoc9rxJ9eZ33mkUrFItPU0pl6IR4IFjGBtpIe43g+xHhPM0/bsdFCt2Q07B1omTkFqUKA1qH71hxFJQio04VipnmPH2kux1IxwOqv0E0VjTCZ4RHuGChxS5aV56hk6McoQBZE0T2iUq783UhwqNQ19M5llVIteJv7n9RIdXHopE3GiqSDzQ0HCkY5QVgEaMkmJ5lNDMJHMZEVkjCUm2hRVMSU4i19eJu5Z46rh3J3XmrdFG2U4gmOogwMX0IQbaIELBCQ8wyu8WU/Wi/VufcxHS1axcwh/YH3+AJi2kiU=</latexit>
L<latexit sha1_base64="UIzmQdPm4e6dU3uEnO6VqVUUQ+8=">AAAB8XicbVBNSwMxFMzWr1q/qh69BIvgqeyKoN4KXkQ8VHBtYbuUbJptQ7PJkrwVytKf4cWDilf/jTf/jdl2D9o6EBhm3iPzJkoFN+C6305lZXVtfaO6Wdva3tndq+8fPBqVacp8qoTS3YgYJrhkPnAQrJtqRpJIsE40vi78zhPThiv5AJOUhQkZSh5zSsBKQS8hMKJE5HfTfr3hNt0Z8DLxStJAJdr9+ldvoGiWMAlUEGMCz00hzIkGTgWb1nqZYSmhYzJkgaWSJMyE+SzyFJ9YZYBjpe2TgGfq742cJMZMkshOFhHNoleI/3lBBvFlmHOZZsAknX8UZwKDwsX9eMA1oyAmlhCquc2K6YhoQsG2VLMleIsnLxP/rHnV9O7PG63bso0qOkLH6BR56AK10A1qIx9RpNAzekVvDjgvzrvzMR+tOOXOIfoD5/MH7biRPg==</latexit><latexit sha1_base64="UIzmQdPm4e6dU3uEnO6VqVUUQ+8=">AAAB8XicbVBNSwMxFMzWr1q/qh69BIvgqeyKoN4KXkQ8VHBtYbuUbJptQ7PJkrwVytKf4cWDilf/jTf/jdl2D9o6EBhm3iPzJkoFN+C6305lZXVtfaO6Wdva3tndq+8fPBqVacp8qoTS3YgYJrhkPnAQrJtqRpJIsE40vi78zhPThiv5AJOUhQkZSh5zSsBKQS8hMKJE5HfTfr3hNt0Z8DLxStJAJdr9+ldvoGiWMAlUEGMCz00hzIkGTgWb1nqZYSmhYzJkgaWSJMyE+SzyFJ9YZYBjpe2TgGfq742cJMZMkshOFhHNoleI/3lBBvFlmHOZZsAknX8UZwKDwsX9eMA1oyAmlhCquc2K6YhoQsG2VLMleIsnLxP/rHnV9O7PG63bso0qOkLH6BR56AK10A1qIx9RpNAzekVvDjgvzrvzMR+tOOXOIfoD5/MH7biRPg==</latexit><latexit sha1_base64="UIzmQdPm4e6dU3uEnO6VqVUUQ+8=">AAAB8XicbVBNSwMxFMzWr1q/qh69BIvgqeyKoN4KXkQ8VHBtYbuUbJptQ7PJkrwVytKf4cWDilf/jTf/jdl2D9o6EBhm3iPzJkoFN+C6305lZXVtfaO6Wdva3tndq+8fPBqVacp8qoTS3YgYJrhkPnAQrJtqRpJIsE40vi78zhPThiv5AJOUhQkZSh5zSsBKQS8hMKJE5HfTfr3hNt0Z8DLxStJAJdr9+ldvoGiWMAlUEGMCz00hzIkGTgWb1nqZYSmhYzJkgaWSJMyE+SzyFJ9YZYBjpe2TgGfq742cJMZMkshOFhHNoleI/3lBBvFlmHOZZsAknX8UZwKDwsX9eMA1oyAmlhCquc2K6YhoQsG2VLMleIsnLxP/rHnV9O7PG63bso0qOkLH6BR56AK10A1qIx9RpNAzekVvDjgvzrvzMR+tOOXOIfoD5/MH7biRPg==</latexit><latexit sha1_base64="UIzmQdPm4e6dU3uEnO6VqVUUQ+8=">AAAB8XicbVBNSwMxFMzWr1q/qh69BIvgqeyKoN4KXkQ8VHBtYbuUbJptQ7PJkrwVytKf4cWDilf/jTf/jdl2D9o6EBhm3iPzJkoFN+C6305lZXVtfaO6Wdva3tndq+8fPBqVacp8qoTS3YgYJrhkPnAQrJtqRpJIsE40vi78zhPThiv5AJOUhQkZSh5zSsBKQS8hMKJE5HfTfr3hNt0Z8DLxStJAJdr9+ldvoGiWMAlUEGMCz00hzIkGTgWb1nqZYSmhYzJkgaWSJMyE+SzyFJ9YZYBjpe2TgGfq742cJMZMkshOFhHNoleI/3lBBvFlmHOZZsAknX8UZwKDwsX9eMA1oyAmlhCquc2K6YhoQsG2VLMleIsnLxP/rHnV9O7PG63bso0qOkLH6BR56AK10A1qIx9RpNAzekVvDjgvzrvzMR+tOOXOIfoD5/MH7biRPg==</latexit>
T = {L(e1, e2, . . . , eK), x1 ⇠ q(x), xt+1 ⇠ q(x|xt, at), H}<latexit sha1_base64="ZL5RnVdScnbZacaHZ6rzokxi/dU=">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</latexit><latexit sha1_base64="ZL5RnVdScnbZacaHZ6rzokxi/dU=">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</latexit><latexit sha1_base64="ZL5RnVdScnbZacaHZ6rzokxi/dU=">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</latexit><latexit sha1_base64="ZL5RnVdScnbZacaHZ6rzokxi/dU=">AAACOHicbVDLahtBEJx1HlaUh2XnmEsTEVCIELvC4ORgEPgi4hxkkCKBViyzo5Y0aPbhmV4jsdFv+ZK/yC2Qiw92yDVf4NGDkEgpGKiu6qanK0yVNOS63529Bw8fPd4vPCk+ffb8xUHp8OizSTItsCMSleheyA0qGWOHJCnspRp5FCrshtOzpd+9Qm1kErdpnuIg4uNYjqTgZKWg1GrDKfg5fKpg4FUBg3oV/GFCZsnP31ZhFnjgGxnBZWW2KnN65y3+SF+sQlXgAVmzCf4iKJXdmrsC7BJvQ8psg1ZQ+mb3iSzCmITixvQ9N6VBzjVJoXBR9DODKRdTPsa+pTGP0Azy1eULeGOVIYwSbV9MsFL/nsh5ZMw8Cm1nxGlitr2l+D+vn9Ho/SCXcZoRxmK9aJQpoASWMcJQahSk5pZwoaX9K4gJ11yQDbtoQ/C2T94lnXrtQ827OC43Pm7SKLBX7DWrMI+dsAZrshbrMMGu2Q92y+6cr86N89P5tW7dczYzL9k/cH7fAyMzp3E=</latexit>
e = (x1, a1, . . . , xH, aH)<latexit sha1_base64="RTvl4VSKDR1Czf8GOPpJeflQwCo=">AAACDnicbVBNS8MwGE7n15xfVY9egkOZMEYrgnoQBl6GpwnWDbZS0jTbwtK0JKlslP0DL/4VLx5UvHr25r8x3XrQzQcSHp7nfd/kffyYUaks69soLC2vrK4V10sbm1vbO+bu3r2MEoGJgyMWibaPJGGUE0dRxUg7FgSFPiMtf3id+a0HIiSN+J0ax8QNUZ/THsVIackzj7t+xIKUTOAVrIw8uwpRdnWDSMkqHHmNTGiceGbZqllTwEVi56QMcjQ980uPwElIuMIMSdmxrVi5KRKKYkYmpW4iSYzwEPVJR1OOQiLddLrPBB5pJYC9SOjDFZyqvztSFEo5Dn1dGSI1kPNeJv7ndRLVu3BTyuNEEY5nD/USBlUEs3BgQAXBio01QVhQ/VeIB0ggrHSEJR2CPb/yInFOa5c1+/asXL/J0yiCA3AIKsAG56AOGqAJHIDBI3gGr+DNeDJejHfjY1ZaMPKeffAHxucP0DKZkQ==</latexit><latexit sha1_base64="RTvl4VSKDR1Czf8GOPpJeflQwCo=">AAACDnicbVBNS8MwGE7n15xfVY9egkOZMEYrgnoQBl6GpwnWDbZS0jTbwtK0JKlslP0DL/4VLx5UvHr25r8x3XrQzQcSHp7nfd/kffyYUaks69soLC2vrK4V10sbm1vbO+bu3r2MEoGJgyMWibaPJGGUE0dRxUg7FgSFPiMtf3id+a0HIiSN+J0ax8QNUZ/THsVIackzj7t+xIKUTOAVrIw8uwpRdnWDSMkqHHmNTGiceGbZqllTwEVi56QMcjQ980uPwElIuMIMSdmxrVi5KRKKYkYmpW4iSYzwEPVJR1OOQiLddLrPBB5pJYC9SOjDFZyqvztSFEo5Dn1dGSI1kPNeJv7ndRLVu3BTyuNEEY5nD/USBlUEs3BgQAXBio01QVhQ/VeIB0ggrHSEJR2CPb/yInFOa5c1+/asXL/J0yiCA3AIKsAG56AOGqAJHIDBI3gGr+DNeDJejHfjY1ZaMPKeffAHxucP0DKZkQ==</latexit><latexit sha1_base64="RTvl4VSKDR1Czf8GOPpJeflQwCo=">AAACDnicbVBNS8MwGE7n15xfVY9egkOZMEYrgnoQBl6GpwnWDbZS0jTbwtK0JKlslP0DL/4VLx5UvHr25r8x3XrQzQcSHp7nfd/kffyYUaks69soLC2vrK4V10sbm1vbO+bu3r2MEoGJgyMWibaPJGGUE0dRxUg7FgSFPiMtf3id+a0HIiSN+J0ax8QNUZ/THsVIackzj7t+xIKUTOAVrIw8uwpRdnWDSMkqHHmNTGiceGbZqllTwEVi56QMcjQ980uPwElIuMIMSdmxrVi5KRKKYkYmpW4iSYzwEPVJR1OOQiLddLrPBB5pJYC9SOjDFZyqvztSFEo5Dn1dGSI1kPNeJv7ndRLVu3BTyuNEEY5nD/USBlUEs3BgQAXBio01QVhQ/VeIB0ggrHSEJR2CPb/yInFOa5c1+/asXL/J0yiCA3AIKsAG56AOGqAJHIDBI3gGr+DNeDJejHfjY1ZaMPKeffAHxucP0DKZkQ==</latexit><latexit sha1_base64="RTvl4VSKDR1Czf8GOPpJeflQwCo=">AAACDnicbVBNS8MwGE7n15xfVY9egkOZMEYrgnoQBl6GpwnWDbZS0jTbwtK0JKlslP0DL/4VLx5UvHr25r8x3XrQzQcSHp7nfd/kffyYUaks69soLC2vrK4V10sbm1vbO+bu3r2MEoGJgyMWibaPJGGUE0dRxUg7FgSFPiMtf3id+a0HIiSN+J0ax8QNUZ/THsVIackzj7t+xIKUTOAVrIw8uwpRdnWDSMkqHHmNTGiceGbZqllTwEVi56QMcjQ980uPwElIuMIMSdmxrVi5KRKKYkYmpW4iSYzwEPVJR1OOQiLddLrPBB5pJYC9SOjDFZyqvztSFEo5Dn1dGSI1kPNeJv7ndRLVu3BTyuNEEY5nD/USBlUEs3BgQAXBio01QVhQ/VeIB0ggrHSEJR2CPb/yInFOa5c1+/asXL/J0yiCA3AIKsAG56AOGqAJHIDBI3gGr+DNeDJejHfjY1ZaMPKeffAHxucP0DKZkQ==</latexit>
Reinforcement Learning
4. Approach
Experiment on Few-Shot Classification
• Omniglot (Lake et al., 2012)

• 50 different alphabets, 1623 characters.

• 20 instances for each characters were drawn by 20 different people.

• 1200 for training, 423 for test.

• Mini-Imagenet (Ravi & Larochelle, 2017)

• Classes for each set: train=64, validation=12, test=24.

5. Experiment
• MAML outperforms state-of-arts algorithms.
Results of Few-Shot Classification
5. Experiment
Effect of 1st Order Approximation
• In 1st order approximation, computation is roughly 33% better.

• Performances are similar.
5. Experiment
Effect of 1st Order Approximation
• In 1st order approximation, computation is roughly 33% better.

• Performances are similar.
5. Experiment
Performances are not that different.
Effect of 1st Order Approximation
• In 1st order approximation, computation is roughly 33% better.

• Performances are similar.
5. Experiment
Some performance is even better.
Experiments on Regression
• Sinusoid function:

• Amplitude (A) and phase (ɸ) are varied between tasks

• A in [0.1, 0.5]

• ɸ in [0, π]

• x in [-5.0, 5.0]

• Loss function: Mean Squared Error (MSE)

• Regressor: 2 hidden layers with 40 units and ReLU

• Training

• Use only 1 gradient step for learner

• K = 5 or 10 example (5-shot learning or 10-shot learning)

• Fixed step size (ɑ=0.01) for Adam optimizer.
5. Experiment
Results of 10-Shot Learning Regression
5. Experiment
[MAML] [Pretraining + Fine-tuning]
The red line is ground truth.

Fit this sine function with only few (10) samples.
Results of 10-Shot Learning Regression
5. Experiment
Above plots are the pre-trained function of two models.

(The prediction of meta-parameter of MAML,

The prediction of co-learned parameter of vanilla multi-task learning)
[MAML] [Pretraining + Fine-tuning]
Results of 10-Shot Learning Regression
5. Experiment
After 1 gradient step update.
[MAML] [Pretraining + Fine-tuning]
Results of 10-Shot Learning Regression
5. Experiment
[MAML] [Pretraining + Fine-tuning]
After 10 gradient step update.
Results of 10-Shot Learning Regression
5. Experiment
* step size 0.02
In the case of MAML, they quickly adapted to new samples.

But, adaptation failed in the case of pretraining model.
[MAML] [Pretraining + Fine-tuning]
Results of 5-Shot Learning Regression
5. Experiment
* step size 0.01
[MAML] [Pretraining + Fine-tuning]
In the 5-shot learning, the difference is pervasive.

Results of 5-Shot Learning Regression
5. Experiment
* step size 0.01
In the 5-shot learning, the difference is pervasive.

Good predictions are also made for ranges not particularly seen.
[MAML] [Pretraining + Fine-tuning]
MAML Needs Only One Gradient Step
5. Experiment
Vanilla pretrained model adapted slowly,

but, the MAML method quickly adapted even in one gradient step.
MAML Needs Only One Gradient Step
5. Experiment
Vanilla pretrained model adapted slowly,

but, the MAML method quickly adapted even in one gradient step.

Meta parameter theta of MAML is more sensitive

than pre-updated theta of vanilla pretrained model.
MSE of the Meta-Parameters
• The performance of the-meta
parameters was not improved
much in training.

• However, the performance of 

the single gradient updated
parameters 

started on meta-parameters

improved as training progressed.
5. Experiment
Experiments on Reinforcement Learning
• rllab benchmark suite

• Neural network policy with two hidden layers of size 100 with ReLU

• Gradients updates are computed using vanilla policy gradient
(REINFORCE) 

and trust-region policy (TRPO) optimization as meta-optimizer.

• Comparison

• Pretraining one policy on all of the tasks and fine-tuning

• Training a policy from randomly initialized weights

• Oracle policy
5. Experiment
Results on Reinforcement Learning
• 2d navigation
5. Experiment
Results on Reinforcement Learning
• Locomotion

• High-dimensional locomotion tasks with the MuJoCo simulator
5. Experiment
Why 1st Order Approximation Works?
6. Discussion
Why 1st Order Approximation Works?
• [Montufar14] The objective function of ReLU network is locally linear.

• So as the other piecewise linear activations (maxout, leaky relu, ...) are
used. 

• Because, any composite function with linear or piecewise linear
functions should be locally linear.

•
6. Discussion
magnification
Solid: Shallow network with 20 hidden units

Dotted: deep network (2 hidden layers) with 10 hidden units at each layer
Why 1st Order Approximation Works?
• [Montufar14] The objective function of ReLU network is locally linear.

• So as the other piecewise linear activations (maxout, leaky relu, ...) are
used. 

• Because, any composite function with linear or piecewise linear
functions should be locally linear.

• Hence, the Hessian matrix is almost zero.

= Hessian matrix is meaningless.
6. Discussion
magnification
Solid: Shallow network with 20 hidden units

Dotted: deep network (2 hidden layers) with 10 hidden units at each layer
Why 1st Order Approximation Works?
• If we change the activation function from ReLU to sigmoid?
6. Discussion
Why 1st Order Approximation Works?
• If we change the activation function from ReLU to sigmoid?
6. Discussion
ReLU sigmoid
w/o 1st order approx. 0.1183 0.1579
w 1st order approx. 0.1183 0.1931
ReLU sigmoid
w/o 1st order approx. 61 96
w 1st order approx. 53 (13% better) 52 (46% better)
[ MSE Comparison ]
[ Learning Time Comparison (minutes) ]
Why the Authors Focused on Deep Networks?
• The authors said their model (MAML) is model-agnostic.

• But, why they focused on deep networks? 

• If it is not deep networks, what kind of problem can occur?
6. Discussion
Why the Authors Focused on Deep Networks?
• Simple linear regression model (y = θx) can be used in the MAML.

• Parametrized, differentiable loss function.
6. Discussion
Why the Authors Focused on Deep Networks?
• Simple linear regression model (y = θx) can be used in the MAML.

• Parametrized, differentiable loss function.

• Let
6. Discussion
Why the Authors Focused on Deep Networks?
• Simple linear regression model (y = θx) can be used in the MAML.

• Parametrized, differentiable loss function.

• Let

• Then, meta parameter θ will be 1.25.
6. Discussion
Why the Authors Focused on Deep Networks?
• Let’s extend the setting. 

• The sample is drawn from for the task .

• Then the meta parameter .
6. Discussion
Why the Authors Focused on Deep Networks?
• Let’s extend the setting. 

• The sample is drawn from for the task .

• Then the meta parameter .

• The MAML use fixed learning rate. 

• More than 2 gradient steps are needed when is far from .
6. Discussion
Why the Authors Focused on Deep Networks?
• Let’s extend the setting. 

• The sample is drawn from for the task .

• Then the meta parameter .

• The MAML use fixed learning rate. 

• More than 2 gradient steps are needed when is far from .

• However, because the dimension of the deep model is large, 

there may be a meta parameter

that can quickly adapt to any task parameter .

• This is why we think the authors focused on the deep models.
6. Discussion
• However, using the deep model doesn’t eliminate the problem
completely.

• The authors mentioned about the generalization problem for
extrapolated tasks.

• The closer the task is to out-of-distribution, the worse the performance.

• How can we solve this problem?
Why the Authors Focused on Deep Networks?
6. Discussion
Update Meta Parameters with N-steps
• The MAML updates using only one gradient step 

when updating the meta parameter.

• The authors say that you can update meta parameters using n
gradient steps.

6. Discussion
Update Meta Parameters with N-steps
• Will it show better performance to update with N gradient steps?

• If so, why the authors didn’t use this method in experiments?
6. Discussion
Update Meta Parameters with N-steps
• Will it show better performance to update with N gradient steps?

• If so, why the authors didn’t use this method in experiments?

• If we apply this method to MAML, the MAML model will find a
sensitive meta parameter in the N-step update.
6. Discussion
Update Meta Parameters with N-steps
• Will it show better performance to update with N gradient steps?

• If so, why the authors didn’t use this method in experiments?

• If we apply this method to MAML, the MAML model will find a
sensitive meta parameter in the N-step update.

• Also, we should calculate (n + 1)-order partial derivatives.

• (c.f.) Hessian matrix is 2-order partial derivatives.

• The computation of (n + 1)-order partial derivatives is
computationally expensive.
6. Discussion
Update Meta Parameters with N-steps
• How about using 1st order approximation with ReLU activation 

and updating meta parameter with N-steps?
6. Discussion
Update Meta Parameters with N-steps
• N-step(s) updates with 1st order approximation.

• Of course, we use ReLU for activation function.
6. Discussion
1-step 2-steps 3-steps
Learning Time 53 73 92
Update Meta Parameters with N-steps
• N-step(s) updates with 1st order approximation.

• Of course, we use ReLU for activation function.
6. Discussion
1-step 2-steps 3-steps
Learning Time 53 73 92
the performance of the meta-parameter 

getting worser when n grows.
Update Meta Parameters with N-steps
• N-step(s) updates with 1st order approximation.

• Of course, we use ReLU for activation function.
6. Discussion
1-step 2-steps 3-steps
Learning Time 53 73 92
Final performance may be unrelated to N.

(More experiments are needed for confirming this issue.)
Unsupervised Learning
• The authors said that the MAML is model agnostic. 

• They showed the MAML is worked on supervised learning.

• How about unsupervised learning?
6. Discussion
• The authors said that the MAML is model agnostic. 

• They showed the MAML is worked on supervised learning.

• How about unsupervised learning?

• We can use autoencoder.

• Autoencoder can make unsupervised learning

be regarded as supervised learning.
Unsupervised Learning
6. Discussion
• MAML is really slow algorithm even with 1st order approximation.

• How can we make the MAML framework be faster?
How to Make the MAML be faster?
6. Discussion
• MAML is really slow algorithm even with 1st order approximation.

• How can we make the MAML framework be faster?

• [Zhenguo 2017] suggests 

optimizing learning rate alpha.

• Instead of putting alpha 

as a hyperparameter, 

train the vector alpha 

as the parameters

with other parameters.

• It outperforms the vanilla MAML.
How to Make the MAML be faster?
6. Discussion
Application
• The MAML is good for a few-shot learning problem.

• Can you guess any applications of few-shot learning problem?
6. Discussion
Application
• The MAML is good for a few-shot learning problem.

• Can you guess any applications of few-shot learning problem?

• We think one of the most effective few-shot learning problem is 

real estate price prediction.

• There are a lot of transactions related to big apartments.

• But, there are many transactions in small apartments.

So, it is difficult to predict.

• Using MAML seems to solve the problem of insufficient sample size of
small apartments.

• Also, other problems would be solved by MAML.
6. Discussion
Thank you!
Reference
• Finn, Chelsea, Pieter Abbeel, and Sergey Levine. "Model-agnostic
meta-learning for fast adaptation of deep networks." arXiv preprint
arXiv:1703.03400 (2017).

• Montufar, Guido F., et al. "On the number of linear regions of deep
neural networks." Advances in neural information processing
systems. 2014.

• Baldi, Pierre. "Autoencoders, unsupervised learning, and deep
architectures." Proceedings of ICML workshop on unsupervised and
transfer learning. 2012.

• Li, Zhenguo, et al. "Meta-sgd: Learning to learn quickly for few shot
learning." arXiv preprint arXiv:1707.09835 (2017)
Reference
• https://www.slideshare.net/TaesuKim3/pr12094-modelagnostic-
metalearning-for-fast-adaptation-of-deep-networks 

• https://people.eecs.berkeley.edu/~cbfinn/_files/
nips2017_metaworkshop.pdf 

• https://github.com/cbfinn/maml

• Lab seminar slide by Young-ki Hong.

More Related Content

What's hot

Life-long / Incremental Learning (DLAI D6L1 2017 UPC Deep Learning for Artifi...
Life-long / Incremental Learning (DLAI D6L1 2017 UPC Deep Learning for Artifi...Life-long / Incremental Learning (DLAI D6L1 2017 UPC Deep Learning for Artifi...
Life-long / Incremental Learning (DLAI D6L1 2017 UPC Deep Learning for Artifi...
Universitat Politècnica de Catalunya
 
Meta learning with memory augmented neural network
Meta learning with memory augmented neural networkMeta learning with memory augmented neural network
Meta learning with memory augmented neural network
Katy Lee
 
Word_Embedding.pptx
Word_Embedding.pptxWord_Embedding.pptx
Word_Embedding.pptx
NameetDaga1
 
META-LEARNING.pptx
META-LEARNING.pptxMETA-LEARNING.pptx
META-LEARNING.pptx
AyanaRukasar
 
Continual Learning with Deep Architectures - Tutorial ICML 2021
Continual Learning with Deep Architectures - Tutorial ICML 2021Continual Learning with Deep Architectures - Tutorial ICML 2021
Continual Learning with Deep Architectures - Tutorial ICML 2021
Vincenzo Lomonaco
 
Bayesian Model-Agnostic Meta-Learning
Bayesian Model-Agnostic Meta-LearningBayesian Model-Agnostic Meta-Learning
Bayesian Model-Agnostic Meta-Learning
Sangwoo Mo
 
Activation functions and Training Algorithms for Deep Neural network
Activation functions and Training Algorithms for Deep Neural networkActivation functions and Training Algorithms for Deep Neural network
Activation functions and Training Algorithms for Deep Neural network
Gayatri Khanvilkar
 
Comparing Incremental Learning Strategies for Convolutional Neural Networks
Comparing Incremental Learning Strategies for Convolutional Neural NetworksComparing Incremental Learning Strategies for Convolutional Neural Networks
Comparing Incremental Learning Strategies for Convolutional Neural Networks
Vincenzo Lomonaco
 
Meta-Learning with Memory-Augmented Neural Networks (MANN)
Meta-Learning with Memory-Augmented Neural Networks (MANN)Meta-Learning with Memory-Augmented Neural Networks (MANN)
Meta-Learning with Memory-Augmented Neural Networks (MANN)
Yeonsu Kim
 
PR-317: MLP-Mixer: An all-MLP Architecture for Vision
PR-317: MLP-Mixer: An all-MLP Architecture for VisionPR-317: MLP-Mixer: An all-MLP Architecture for Vision
PR-317: MLP-Mixer: An all-MLP Architecture for Vision
Jinwon Lee
 
Introduction to continual learning
Introduction to continual learningIntroduction to continual learning
Introduction to continual learning
Nguyen Giang
 
Matching networks for one shot learning
Matching networks for one shot learningMatching networks for one shot learning
Matching networks for one shot learning
Kazuki Fujikawa
 
Overcoming catastrophic forgetting in neural network
Overcoming catastrophic forgetting in neural networkOvercoming catastrophic forgetting in neural network
Overcoming catastrophic forgetting in neural network
Katy Lee
 
Perceptron
PerceptronPerceptron
Perceptron
Nagarajan
 
Attention Is All You Need
Attention Is All You NeedAttention Is All You Need
Attention Is All You Need
Illia Polosukhin
 
Meta-learning and the ELBO
Meta-learning and the ELBOMeta-learning and the ELBO
Meta-learning and the ELBO
Yoonho Lee
 
[Paper review] BERT
[Paper review] BERT[Paper review] BERT
[Paper review] BERT
JEE HYUN PARK
 
Intro to modelling-supervised learning
Intro to modelling-supervised learningIntro to modelling-supervised learning
Intro to modelling-supervised learning
Justin Sebok
 
What is the Expectation Maximization (EM) Algorithm?
What is the Expectation Maximization (EM) Algorithm?What is the Expectation Maximization (EM) Algorithm?
What is the Expectation Maximization (EM) Algorithm?
Kazuki Yoshida
 
論文紹介「A Perspective View and Survey of Meta-Learning」
論文紹介「A Perspective View and Survey of Meta-Learning」論文紹介「A Perspective View and Survey of Meta-Learning」
論文紹介「A Perspective View and Survey of Meta-Learning」
Kota Matsui
 

What's hot (20)

Life-long / Incremental Learning (DLAI D6L1 2017 UPC Deep Learning for Artifi...
Life-long / Incremental Learning (DLAI D6L1 2017 UPC Deep Learning for Artifi...Life-long / Incremental Learning (DLAI D6L1 2017 UPC Deep Learning for Artifi...
Life-long / Incremental Learning (DLAI D6L1 2017 UPC Deep Learning for Artifi...
 
Meta learning with memory augmented neural network
Meta learning with memory augmented neural networkMeta learning with memory augmented neural network
Meta learning with memory augmented neural network
 
Word_Embedding.pptx
Word_Embedding.pptxWord_Embedding.pptx
Word_Embedding.pptx
 
META-LEARNING.pptx
META-LEARNING.pptxMETA-LEARNING.pptx
META-LEARNING.pptx
 
Continual Learning with Deep Architectures - Tutorial ICML 2021
Continual Learning with Deep Architectures - Tutorial ICML 2021Continual Learning with Deep Architectures - Tutorial ICML 2021
Continual Learning with Deep Architectures - Tutorial ICML 2021
 
Bayesian Model-Agnostic Meta-Learning
Bayesian Model-Agnostic Meta-LearningBayesian Model-Agnostic Meta-Learning
Bayesian Model-Agnostic Meta-Learning
 
Activation functions and Training Algorithms for Deep Neural network
Activation functions and Training Algorithms for Deep Neural networkActivation functions and Training Algorithms for Deep Neural network
Activation functions and Training Algorithms for Deep Neural network
 
Comparing Incremental Learning Strategies for Convolutional Neural Networks
Comparing Incremental Learning Strategies for Convolutional Neural NetworksComparing Incremental Learning Strategies for Convolutional Neural Networks
Comparing Incremental Learning Strategies for Convolutional Neural Networks
 
Meta-Learning with Memory-Augmented Neural Networks (MANN)
Meta-Learning with Memory-Augmented Neural Networks (MANN)Meta-Learning with Memory-Augmented Neural Networks (MANN)
Meta-Learning with Memory-Augmented Neural Networks (MANN)
 
PR-317: MLP-Mixer: An all-MLP Architecture for Vision
PR-317: MLP-Mixer: An all-MLP Architecture for VisionPR-317: MLP-Mixer: An all-MLP Architecture for Vision
PR-317: MLP-Mixer: An all-MLP Architecture for Vision
 
Introduction to continual learning
Introduction to continual learningIntroduction to continual learning
Introduction to continual learning
 
Matching networks for one shot learning
Matching networks for one shot learningMatching networks for one shot learning
Matching networks for one shot learning
 
Overcoming catastrophic forgetting in neural network
Overcoming catastrophic forgetting in neural networkOvercoming catastrophic forgetting in neural network
Overcoming catastrophic forgetting in neural network
 
Perceptron
PerceptronPerceptron
Perceptron
 
Attention Is All You Need
Attention Is All You NeedAttention Is All You Need
Attention Is All You Need
 
Meta-learning and the ELBO
Meta-learning and the ELBOMeta-learning and the ELBO
Meta-learning and the ELBO
 
[Paper review] BERT
[Paper review] BERT[Paper review] BERT
[Paper review] BERT
 
Intro to modelling-supervised learning
Intro to modelling-supervised learningIntro to modelling-supervised learning
Intro to modelling-supervised learning
 
What is the Expectation Maximization (EM) Algorithm?
What is the Expectation Maximization (EM) Algorithm?What is the Expectation Maximization (EM) Algorithm?
What is the Expectation Maximization (EM) Algorithm?
 
論文紹介「A Perspective View and Survey of Meta-Learning」
論文紹介「A Perspective View and Survey of Meta-Learning」論文紹介「A Perspective View and Survey of Meta-Learning」
論文紹介「A Perspective View and Survey of Meta-Learning」
 

Similar to Introduction to MAML (Model Agnostic Meta Learning) with Discussions

Optimization as a model for few shot learning
Optimization as a model for few shot learningOptimization as a model for few shot learning
Optimization as a model for few shot learning
Katy Lee
 
AI_Unit-4_Learning.pptx
AI_Unit-4_Learning.pptxAI_Unit-4_Learning.pptx
AI_Unit-4_Learning.pptx
MohammadAsim91
 
Experimental Design for Distributed Machine Learning with Myles Baker
Experimental Design for Distributed Machine Learning with Myles BakerExperimental Design for Distributed Machine Learning with Myles Baker
Experimental Design for Distributed Machine Learning with Myles Baker
Databricks
 
Efficient Off-Policy Meta-Reinforcement Learning via Probabilistic Context Va...
Efficient Off-Policy Meta-Reinforcement Learning via Probabilistic Context Va...Efficient Off-Policy Meta-Reinforcement Learning via Probabilistic Context Va...
Efficient Off-Policy Meta-Reinforcement Learning via Probabilistic Context Va...
Dongmin Lee
 
Learning to Learn by Gradient Descent by Gradient Descent
Learning to Learn by Gradient Descent by Gradient DescentLearning to Learn by Gradient Descent by Gradient Descent
Learning to Learn by Gradient Descent by Gradient Descent
Katy Lee
 
How can pre-training help to solve the cold start problem?
How can pre-training help to solve the cold start problem?How can pre-training help to solve the cold start problem?
How can pre-training help to solve the cold start problem?Lokesh Vadlamudi
 
Hyperparameter Tuning
Hyperparameter TuningHyperparameter Tuning
Hyperparameter Tuning
Jon Lederman
 
Multi Task Learning and Meta Learning
Multi Task Learning and Meta LearningMulti Task Learning and Meta Learning
Multi Task Learning and Meta Learning
Srilalitha Veerubhotla
 
Learning how to learn
Learning how to learnLearning how to learn
Learning how to learn
Joaquin Vanschoren
 
NPTL - Machine Learning by Madhur Jatiya.pdf
NPTL - Machine Learning by Madhur Jatiya.pdfNPTL - Machine Learning by Madhur Jatiya.pdf
NPTL - Machine Learning by Madhur Jatiya.pdf
Mr. Moms
 
Understanding Mahout classification documentation
Understanding Mahout  classification documentationUnderstanding Mahout  classification documentation
Understanding Mahout classification documentation
Naveen Kumar
 
deeplearning 67589.pptx
deeplearning 67589.pptxdeeplearning 67589.pptx
deeplearning 67589.pptx
CoolDude357501
 
Summer Report on Mathematics for Machine learning: Imperial College of London
Summer Report on Mathematics for Machine learning: Imperial College of LondonSummer Report on Mathematics for Machine learning: Imperial College of London
Summer Report on Mathematics for Machine learning: Imperial College of London
Yash Khanna
 
Module III MachineLearningSparkML.pptx
Module III MachineLearningSparkML.pptxModule III MachineLearningSparkML.pptx
Module III MachineLearningSparkML.pptx
Rahul Borate
 
Endsem AI merged.pdf
Endsem AI merged.pdfEndsem AI merged.pdf
Endsem AI merged.pdf
ShivamMishra603376
 
a deep reinforced model for abstractive summarization
a deep reinforced model for abstractive summarizationa deep reinforced model for abstractive summarization
a deep reinforced model for abstractive summarization
JEE HYUN PARK
 
Machine learning
Machine learningMachine learning
Machine learning
deepakbagam
 
supervised and unsupervised learning
supervised and unsupervised learningsupervised and unsupervised learning
supervised and unsupervised learning
shivani saluja
 
How to test models using php unit testing framework?
How to test models using php unit testing framework?How to test models using php unit testing framework?
How to test models using php unit testing framework?
satejsahu
 
An explanation of machine learning for business
An explanation of machine learning for businessAn explanation of machine learning for business
An explanation of machine learning for business
Clement Levallois
 

Similar to Introduction to MAML (Model Agnostic Meta Learning) with Discussions (20)

Optimization as a model for few shot learning
Optimization as a model for few shot learningOptimization as a model for few shot learning
Optimization as a model for few shot learning
 
AI_Unit-4_Learning.pptx
AI_Unit-4_Learning.pptxAI_Unit-4_Learning.pptx
AI_Unit-4_Learning.pptx
 
Experimental Design for Distributed Machine Learning with Myles Baker
Experimental Design for Distributed Machine Learning with Myles BakerExperimental Design for Distributed Machine Learning with Myles Baker
Experimental Design for Distributed Machine Learning with Myles Baker
 
Efficient Off-Policy Meta-Reinforcement Learning via Probabilistic Context Va...
Efficient Off-Policy Meta-Reinforcement Learning via Probabilistic Context Va...Efficient Off-Policy Meta-Reinforcement Learning via Probabilistic Context Va...
Efficient Off-Policy Meta-Reinforcement Learning via Probabilistic Context Va...
 
Learning to Learn by Gradient Descent by Gradient Descent
Learning to Learn by Gradient Descent by Gradient DescentLearning to Learn by Gradient Descent by Gradient Descent
Learning to Learn by Gradient Descent by Gradient Descent
 
How can pre-training help to solve the cold start problem?
How can pre-training help to solve the cold start problem?How can pre-training help to solve the cold start problem?
How can pre-training help to solve the cold start problem?
 
Hyperparameter Tuning
Hyperparameter TuningHyperparameter Tuning
Hyperparameter Tuning
 
Multi Task Learning and Meta Learning
Multi Task Learning and Meta LearningMulti Task Learning and Meta Learning
Multi Task Learning and Meta Learning
 
Learning how to learn
Learning how to learnLearning how to learn
Learning how to learn
 
NPTL - Machine Learning by Madhur Jatiya.pdf
NPTL - Machine Learning by Madhur Jatiya.pdfNPTL - Machine Learning by Madhur Jatiya.pdf
NPTL - Machine Learning by Madhur Jatiya.pdf
 
Understanding Mahout classification documentation
Understanding Mahout  classification documentationUnderstanding Mahout  classification documentation
Understanding Mahout classification documentation
 
deeplearning 67589.pptx
deeplearning 67589.pptxdeeplearning 67589.pptx
deeplearning 67589.pptx
 
Summer Report on Mathematics for Machine learning: Imperial College of London
Summer Report on Mathematics for Machine learning: Imperial College of LondonSummer Report on Mathematics for Machine learning: Imperial College of London
Summer Report on Mathematics for Machine learning: Imperial College of London
 
Module III MachineLearningSparkML.pptx
Module III MachineLearningSparkML.pptxModule III MachineLearningSparkML.pptx
Module III MachineLearningSparkML.pptx
 
Endsem AI merged.pdf
Endsem AI merged.pdfEndsem AI merged.pdf
Endsem AI merged.pdf
 
a deep reinforced model for abstractive summarization
a deep reinforced model for abstractive summarizationa deep reinforced model for abstractive summarization
a deep reinforced model for abstractive summarization
 
Machine learning
Machine learningMachine learning
Machine learning
 
supervised and unsupervised learning
supervised and unsupervised learningsupervised and unsupervised learning
supervised and unsupervised learning
 
How to test models using php unit testing framework?
How to test models using php unit testing framework?How to test models using php unit testing framework?
How to test models using php unit testing framework?
 
An explanation of machine learning for business
An explanation of machine learning for businessAn explanation of machine learning for business
An explanation of machine learning for business
 

More from Joonyoung Yi

Mixture-Rank Matrix Approximation for Collaborative Filtering
Mixture-Rank Matrix Approximation for Collaborative FilteringMixture-Rank Matrix Approximation for Collaborative Filtering
Mixture-Rank Matrix Approximation for Collaborative Filtering
Joonyoung Yi
 
Sparsity Normalization: Stabilizing the Expected Outputs of Deep Networks
Sparsity Normalization: Stabilizing the Expected Outputs of Deep NetworksSparsity Normalization: Stabilizing the Expected Outputs of Deep Networks
Sparsity Normalization: Stabilizing the Expected Outputs of Deep Networks
Joonyoung Yi
 
Low-rank Matrix Approximation with Stability
Low-rank Matrix Approximation with StabilityLow-rank Matrix Approximation with Stability
Low-rank Matrix Approximation with Stability
Joonyoung Yi
 
A Neural Autoregressive Approach to Collaborative Filtering (CF-NADE) Slide
A Neural Autoregressive Approach to Collaborative Filtering (CF-NADE) Slide A Neural Autoregressive Approach to Collaborative Filtering (CF-NADE) Slide
A Neural Autoregressive Approach to Collaborative Filtering (CF-NADE) Slide
Joonyoung Yi
 
Introduction to XGBoost
Introduction to XGBoostIntroduction to XGBoost
Introduction to XGBoost
Joonyoung Yi
 
Why biased matrix factorization works well?
Why biased matrix factorization works well?Why biased matrix factorization works well?
Why biased matrix factorization works well?
Joonyoung Yi
 
Dynamically Expandable Network (DEN)
Dynamically Expandable Network (DEN)Dynamically Expandable Network (DEN)
Dynamically Expandable Network (DEN)
Joonyoung Yi
 
Introduction to Low-rank Matrix Completion
Introduction to Low-rank Matrix CompletionIntroduction to Low-rank Matrix Completion
Introduction to Low-rank Matrix Completion
Joonyoung Yi
 
Exact Matrix Completion via Convex Optimization Slide (PPT)
Exact Matrix Completion via Convex Optimization Slide (PPT)Exact Matrix Completion via Convex Optimization Slide (PPT)
Exact Matrix Completion via Convex Optimization Slide (PPT)
Joonyoung Yi
 

More from Joonyoung Yi (9)

Mixture-Rank Matrix Approximation for Collaborative Filtering
Mixture-Rank Matrix Approximation for Collaborative FilteringMixture-Rank Matrix Approximation for Collaborative Filtering
Mixture-Rank Matrix Approximation for Collaborative Filtering
 
Sparsity Normalization: Stabilizing the Expected Outputs of Deep Networks
Sparsity Normalization: Stabilizing the Expected Outputs of Deep NetworksSparsity Normalization: Stabilizing the Expected Outputs of Deep Networks
Sparsity Normalization: Stabilizing the Expected Outputs of Deep Networks
 
Low-rank Matrix Approximation with Stability
Low-rank Matrix Approximation with StabilityLow-rank Matrix Approximation with Stability
Low-rank Matrix Approximation with Stability
 
A Neural Autoregressive Approach to Collaborative Filtering (CF-NADE) Slide
A Neural Autoregressive Approach to Collaborative Filtering (CF-NADE) Slide A Neural Autoregressive Approach to Collaborative Filtering (CF-NADE) Slide
A Neural Autoregressive Approach to Collaborative Filtering (CF-NADE) Slide
 
Introduction to XGBoost
Introduction to XGBoostIntroduction to XGBoost
Introduction to XGBoost
 
Why biased matrix factorization works well?
Why biased matrix factorization works well?Why biased matrix factorization works well?
Why biased matrix factorization works well?
 
Dynamically Expandable Network (DEN)
Dynamically Expandable Network (DEN)Dynamically Expandable Network (DEN)
Dynamically Expandable Network (DEN)
 
Introduction to Low-rank Matrix Completion
Introduction to Low-rank Matrix CompletionIntroduction to Low-rank Matrix Completion
Introduction to Low-rank Matrix Completion
 
Exact Matrix Completion via Convex Optimization Slide (PPT)
Exact Matrix Completion via Convex Optimization Slide (PPT)Exact Matrix Completion via Convex Optimization Slide (PPT)
Exact Matrix Completion via Convex Optimization Slide (PPT)
 

Recently uploaded

Sample_Global Non-invasive Prenatal Testing (NIPT) Market, 2019-2030.pdf
Sample_Global Non-invasive Prenatal Testing (NIPT) Market, 2019-2030.pdfSample_Global Non-invasive Prenatal Testing (NIPT) Market, 2019-2030.pdf
Sample_Global Non-invasive Prenatal Testing (NIPT) Market, 2019-2030.pdf
Linda486226
 
Chatty Kathy - UNC Bootcamp Final Project Presentation - Final Version - 5.23...
Chatty Kathy - UNC Bootcamp Final Project Presentation - Final Version - 5.23...Chatty Kathy - UNC Bootcamp Final Project Presentation - Final Version - 5.23...
Chatty Kathy - UNC Bootcamp Final Project Presentation - Final Version - 5.23...
John Andrews
 
Q1’2024 Update: MYCI’s Leap Year Rebound
Q1’2024 Update: MYCI’s Leap Year ReboundQ1’2024 Update: MYCI’s Leap Year Rebound
Q1’2024 Update: MYCI’s Leap Year Rebound
Oppotus
 
一比一原版(NYU毕业证)纽约大学毕业证成绩单
一比一原版(NYU毕业证)纽约大学毕业证成绩单一比一原版(NYU毕业证)纽约大学毕业证成绩单
一比一原版(NYU毕业证)纽约大学毕业证成绩单
ewymefz
 
The affect of service quality and online reviews on customer loyalty in the E...
The affect of service quality and online reviews on customer loyalty in the E...The affect of service quality and online reviews on customer loyalty in the E...
The affect of service quality and online reviews on customer loyalty in the E...
jerlynmaetalle
 
一比一原版(UniSA毕业证书)南澳大学毕业证如何办理
一比一原版(UniSA毕业证书)南澳大学毕业证如何办理一比一原版(UniSA毕业证书)南澳大学毕业证如何办理
一比一原版(UniSA毕业证书)南澳大学毕业证如何办理
slg6lamcq
 
一比一原版(TWU毕业证)西三一大学毕业证成绩单
一比一原版(TWU毕业证)西三一大学毕业证成绩单一比一原版(TWU毕业证)西三一大学毕业证成绩单
一比一原版(TWU毕业证)西三一大学毕业证成绩单
ocavb
 
一比一原版(RUG毕业证)格罗宁根大学毕业证成绩单
一比一原版(RUG毕业证)格罗宁根大学毕业证成绩单一比一原版(RUG毕业证)格罗宁根大学毕业证成绩单
一比一原版(RUG毕业证)格罗宁根大学毕业证成绩单
vcaxypu
 
一比一原版(CBU毕业证)不列颠海角大学毕业证成绩单
一比一原版(CBU毕业证)不列颠海角大学毕业证成绩单一比一原版(CBU毕业证)不列颠海角大学毕业证成绩单
一比一原版(CBU毕业证)不列颠海角大学毕业证成绩单
nscud
 
一比一原版(Adelaide毕业证书)阿德莱德大学毕业证如何办理
一比一原版(Adelaide毕业证书)阿德莱德大学毕业证如何办理一比一原版(Adelaide毕业证书)阿德莱德大学毕业证如何办理
一比一原版(Adelaide毕业证书)阿德莱德大学毕业证如何办理
slg6lamcq
 
一比一原版(YU毕业证)约克大学毕业证成绩单
一比一原版(YU毕业证)约克大学毕业证成绩单一比一原版(YU毕业证)约克大学毕业证成绩单
一比一原版(YU毕业证)约克大学毕业证成绩单
enxupq
 
SOCRadar Germany 2024 Threat Landscape Report
SOCRadar Germany 2024 Threat Landscape ReportSOCRadar Germany 2024 Threat Landscape Report
SOCRadar Germany 2024 Threat Landscape Report
SOCRadar
 
一比一原版(CBU毕业证)卡普顿大学毕业证如何办理
一比一原版(CBU毕业证)卡普顿大学毕业证如何办理一比一原版(CBU毕业证)卡普顿大学毕业证如何办理
一比一原版(CBU毕业证)卡普顿大学毕业证如何办理
ahzuo
 
Malana- Gimlet Market Analysis (Portfolio 2)
Malana- Gimlet Market Analysis (Portfolio 2)Malana- Gimlet Market Analysis (Portfolio 2)
Malana- Gimlet Market Analysis (Portfolio 2)
TravisMalana
 
一比一原版(CU毕业证)卡尔顿大学毕业证成绩单
一比一原版(CU毕业证)卡尔顿大学毕业证成绩单一比一原版(CU毕业证)卡尔顿大学毕业证成绩单
一比一原版(CU毕业证)卡尔顿大学毕业证成绩单
yhkoc
 
Machine learning and optimization techniques for electrical drives.pptx
Machine learning and optimization techniques for electrical drives.pptxMachine learning and optimization techniques for electrical drives.pptx
Machine learning and optimization techniques for electrical drives.pptx
balafet
 
一比一原版(UPenn毕业证)宾夕法尼亚大学毕业证成绩单
一比一原版(UPenn毕业证)宾夕法尼亚大学毕业证成绩单一比一原版(UPenn毕业证)宾夕法尼亚大学毕业证成绩单
一比一原版(UPenn毕业证)宾夕法尼亚大学毕业证成绩单
ewymefz
 
Algorithmic optimizations for Dynamic Levelwise PageRank (from STICD) : SHORT...
Algorithmic optimizations for Dynamic Levelwise PageRank (from STICD) : SHORT...Algorithmic optimizations for Dynamic Levelwise PageRank (from STICD) : SHORT...
Algorithmic optimizations for Dynamic Levelwise PageRank (from STICD) : SHORT...
Subhajit Sahu
 
Predicting Product Ad Campaign Performance: A Data Analysis Project Presentation
Predicting Product Ad Campaign Performance: A Data Analysis Project PresentationPredicting Product Ad Campaign Performance: A Data Analysis Project Presentation
Predicting Product Ad Campaign Performance: A Data Analysis Project Presentation
Boston Institute of Analytics
 
Empowering Data Analytics Ecosystem.pptx
Empowering Data Analytics Ecosystem.pptxEmpowering Data Analytics Ecosystem.pptx
Empowering Data Analytics Ecosystem.pptx
benishzehra469
 

Recently uploaded (20)

Sample_Global Non-invasive Prenatal Testing (NIPT) Market, 2019-2030.pdf
Sample_Global Non-invasive Prenatal Testing (NIPT) Market, 2019-2030.pdfSample_Global Non-invasive Prenatal Testing (NIPT) Market, 2019-2030.pdf
Sample_Global Non-invasive Prenatal Testing (NIPT) Market, 2019-2030.pdf
 
Chatty Kathy - UNC Bootcamp Final Project Presentation - Final Version - 5.23...
Chatty Kathy - UNC Bootcamp Final Project Presentation - Final Version - 5.23...Chatty Kathy - UNC Bootcamp Final Project Presentation - Final Version - 5.23...
Chatty Kathy - UNC Bootcamp Final Project Presentation - Final Version - 5.23...
 
Q1’2024 Update: MYCI’s Leap Year Rebound
Q1’2024 Update: MYCI’s Leap Year ReboundQ1’2024 Update: MYCI’s Leap Year Rebound
Q1’2024 Update: MYCI’s Leap Year Rebound
 
一比一原版(NYU毕业证)纽约大学毕业证成绩单
一比一原版(NYU毕业证)纽约大学毕业证成绩单一比一原版(NYU毕业证)纽约大学毕业证成绩单
一比一原版(NYU毕业证)纽约大学毕业证成绩单
 
The affect of service quality and online reviews on customer loyalty in the E...
The affect of service quality and online reviews on customer loyalty in the E...The affect of service quality and online reviews on customer loyalty in the E...
The affect of service quality and online reviews on customer loyalty in the E...
 
一比一原版(UniSA毕业证书)南澳大学毕业证如何办理
一比一原版(UniSA毕业证书)南澳大学毕业证如何办理一比一原版(UniSA毕业证书)南澳大学毕业证如何办理
一比一原版(UniSA毕业证书)南澳大学毕业证如何办理
 
一比一原版(TWU毕业证)西三一大学毕业证成绩单
一比一原版(TWU毕业证)西三一大学毕业证成绩单一比一原版(TWU毕业证)西三一大学毕业证成绩单
一比一原版(TWU毕业证)西三一大学毕业证成绩单
 
一比一原版(RUG毕业证)格罗宁根大学毕业证成绩单
一比一原版(RUG毕业证)格罗宁根大学毕业证成绩单一比一原版(RUG毕业证)格罗宁根大学毕业证成绩单
一比一原版(RUG毕业证)格罗宁根大学毕业证成绩单
 
一比一原版(CBU毕业证)不列颠海角大学毕业证成绩单
一比一原版(CBU毕业证)不列颠海角大学毕业证成绩单一比一原版(CBU毕业证)不列颠海角大学毕业证成绩单
一比一原版(CBU毕业证)不列颠海角大学毕业证成绩单
 
一比一原版(Adelaide毕业证书)阿德莱德大学毕业证如何办理
一比一原版(Adelaide毕业证书)阿德莱德大学毕业证如何办理一比一原版(Adelaide毕业证书)阿德莱德大学毕业证如何办理
一比一原版(Adelaide毕业证书)阿德莱德大学毕业证如何办理
 
一比一原版(YU毕业证)约克大学毕业证成绩单
一比一原版(YU毕业证)约克大学毕业证成绩单一比一原版(YU毕业证)约克大学毕业证成绩单
一比一原版(YU毕业证)约克大学毕业证成绩单
 
SOCRadar Germany 2024 Threat Landscape Report
SOCRadar Germany 2024 Threat Landscape ReportSOCRadar Germany 2024 Threat Landscape Report
SOCRadar Germany 2024 Threat Landscape Report
 
一比一原版(CBU毕业证)卡普顿大学毕业证如何办理
一比一原版(CBU毕业证)卡普顿大学毕业证如何办理一比一原版(CBU毕业证)卡普顿大学毕业证如何办理
一比一原版(CBU毕业证)卡普顿大学毕业证如何办理
 
Malana- Gimlet Market Analysis (Portfolio 2)
Malana- Gimlet Market Analysis (Portfolio 2)Malana- Gimlet Market Analysis (Portfolio 2)
Malana- Gimlet Market Analysis (Portfolio 2)
 
一比一原版(CU毕业证)卡尔顿大学毕业证成绩单
一比一原版(CU毕业证)卡尔顿大学毕业证成绩单一比一原版(CU毕业证)卡尔顿大学毕业证成绩单
一比一原版(CU毕业证)卡尔顿大学毕业证成绩单
 
Machine learning and optimization techniques for electrical drives.pptx
Machine learning and optimization techniques for electrical drives.pptxMachine learning and optimization techniques for electrical drives.pptx
Machine learning and optimization techniques for electrical drives.pptx
 
一比一原版(UPenn毕业证)宾夕法尼亚大学毕业证成绩单
一比一原版(UPenn毕业证)宾夕法尼亚大学毕业证成绩单一比一原版(UPenn毕业证)宾夕法尼亚大学毕业证成绩单
一比一原版(UPenn毕业证)宾夕法尼亚大学毕业证成绩单
 
Algorithmic optimizations for Dynamic Levelwise PageRank (from STICD) : SHORT...
Algorithmic optimizations for Dynamic Levelwise PageRank (from STICD) : SHORT...Algorithmic optimizations for Dynamic Levelwise PageRank (from STICD) : SHORT...
Algorithmic optimizations for Dynamic Levelwise PageRank (from STICD) : SHORT...
 
Predicting Product Ad Campaign Performance: A Data Analysis Project Presentation
Predicting Product Ad Campaign Performance: A Data Analysis Project PresentationPredicting Product Ad Campaign Performance: A Data Analysis Project Presentation
Predicting Product Ad Campaign Performance: A Data Analysis Project Presentation
 
Empowering Data Analytics Ecosystem.pptx
Empowering Data Analytics Ecosystem.pptxEmpowering Data Analytics Ecosystem.pptx
Empowering Data Analytics Ecosystem.pptx
 

Introduction to MAML (Model Agnostic Meta Learning) with Discussions

  • 1. MAML: Model-Agnostic Meta-Learning for Fast Adaptation for Deep Networks HyeongJoo Hwang Jeonggwan Lee Joonyoung Yi 2018. 11. 28.
  • 2. Contents 1. Meta-Learning Problem Setup (Definition & Goal) 2. Prior Work 3. Model Agnostic Meta Learning 1. Characteristics 2. Intuition 4. Approach 1. Supervised Learning 2. Gradient of Gradient 3. Reinforcement Learning 5. Experiment 6. Discussion
  • 3. Meta-Learning Problem Set-up • Definition of meta-learning: • Learner is trained by meta-learner to be able to learn on many different tasks. • Goal: • Learner quickly learn new tasks from a small amount of new data. 1. Problem Set-up
  • 4. Meta Supervised Learning • Supervised learning: • Meta supervised learning? 1. Problem Set-up
  • 10. Meta Supervised Learning 1. Problem Set-up
  • 11. Meta Supervised Learning 1. Problem Set-up
  • 12. Prior Work • Learning an update function or update rule • LSTM optimizer
 (Learning to learn by gradient descent by gradient descent) • Meta LSTM optimizer
 (Optimization as a model for few-shot learning) • Few shot (or meta) learning for specific tasks • Generative modeling (Neural Statistician) • Image classification (Matching Net., Prototypical Net.) • Reinforcement learning 
 (Benchmarking deep reinforcement learning for continuous control) 2. Prior Work
  • 13. Prior Work • Memory-augmented meta-learning • Memory-augmented models on many tasks
 (Meta-learning with memory-augmented neural networks) • Parameter initialization of deep networks • Exploring sensitivity while maintaining Internal representation
 (Overcoming catastrophic forgetting in NN) • Data-dependent Initializer
 (Data-Dependent Initialization of CNN) • learned initialization
 (Gradient-based hyperparameter optimization through reversible learning) 2. Prior Work
  • 14. Model-Agnostic Meta-Learning (MAML) • Goal: • Quickly adapt to new tasks on distribution 
 with only small amount of data and 
 with only a few gradient steps, 
 even one gradient step. • Learner: • Learn a new task by using a single gradient step. • Meta-learner: • Learn a generalized parameter initialization of model. 3. Model-Agnostic Meta-Learning (MAML)
  • 15. Characteristics of MAML • The MAML learner’s weights are updated using the gradient, 
 rather than a learned update. • Not need additional parameters nor require a particular learner architecture. • Fast adaptability through good parameter initialization • Explicitly optimizes to learn Internal representation
 (i.e. suitable for many tasks). • Maximizes sensitivity of new task losses to the model parameters. 3. Model-Agnostic Meta-Learning (MAML)
  • 16. Characteristics of MAML • Model-agnostic (No matter whatever model is) • Classification & regression with differentiable losses, Policy gradient RL. • The model should be parameterized. • No other assumptions on the form of the model. • Task-agnostic (No matter whatever task is) • Adopted all knowledge-transferable tasks. • No other assumption is required. 3. Model-Agnostic Meta-Learning (MAML)
  • 17. • Some internal representations
 are more transferrable than others.
 • Desired model parameter set is θ such that:
 Applying one (or a small # of) gradient step to θ
 on a new task will produce maximally effective behavior.
 
 → Find θ that commonly decreases loss of each task 
 after adaptation. Intuition of MAML 3. Model-Agnostic Meta-Learning (MAML)
  • 18. Supervised Learning • Notations • Model: (Model-agnostic) • Task distribution: • For each Task: • Data distribution: q (K samples are drawn from q) • Loss function: • MSE for Regression • Cross entropy for classification • Any other differentiable loss functions can be used. 4. Approach f✓<latexit sha1_base64="vYib4ronG48FFWnNXVHMYLu2Hrs=">AAAB7nicbVBNS8NAEJ3Ur1q/qh69LBbBU0lEUG8FL+KpgrGFNpTNdtMu3Wzi7kQooX/CiwcVr/4eb/4bt20O2vpg4PHeDDPzwlQKg6777ZRWVtfWN8qbla3tnd296v7Bg0kyzbjPEpnodkgNl0JxHwVK3k41p3EoeSscXU/91hPXRiTqHscpD2I6UCISjKKV2lGvi0OOtFetuXV3BrJMvILUoECzV/3q9hOWxVwhk9SYjuemGORUo2CSTyrdzPCUshEd8I6lisbcBPns3gk5sUqfRIm2pZDM1N8TOY2NGceh7YwpDs2iNxX/8zoZRpdBLlSaIVdsvijKJMGETJ8nfaE5Qzm2hDIt7K2EDammDG1EFRuCt/jyMvHP6ld17+681rgt0ijDERzDKXhwAQ24gSb4wEDCM7zCm/PovDjvzse8teQUM4fwB87nD4uEj90=</latexit><latexit sha1_base64="vYib4ronG48FFWnNXVHMYLu2Hrs=">AAAB7nicbVBNS8NAEJ3Ur1q/qh69LBbBU0lEUG8FL+KpgrGFNpTNdtMu3Wzi7kQooX/CiwcVr/4eb/4bt20O2vpg4PHeDDPzwlQKg6777ZRWVtfWN8qbla3tnd296v7Bg0kyzbjPEpnodkgNl0JxHwVK3k41p3EoeSscXU/91hPXRiTqHscpD2I6UCISjKKV2lGvi0OOtFetuXV3BrJMvILUoECzV/3q9hOWxVwhk9SYjuemGORUo2CSTyrdzPCUshEd8I6lisbcBPns3gk5sUqfRIm2pZDM1N8TOY2NGceh7YwpDs2iNxX/8zoZRpdBLlSaIVdsvijKJMGETJ8nfaE5Qzm2hDIt7K2EDammDG1EFRuCt/jyMvHP6ld17+681rgt0ijDERzDKXhwAQ24gSb4wEDCM7zCm/PovDjvzse8teQUM4fwB87nD4uEj90=</latexit><latexit sha1_base64="vYib4ronG48FFWnNXVHMYLu2Hrs=">AAAB7nicbVBNS8NAEJ3Ur1q/qh69LBbBU0lEUG8FL+KpgrGFNpTNdtMu3Wzi7kQooX/CiwcVr/4eb/4bt20O2vpg4PHeDDPzwlQKg6777ZRWVtfWN8qbla3tnd296v7Bg0kyzbjPEpnodkgNl0JxHwVK3k41p3EoeSscXU/91hPXRiTqHscpD2I6UCISjKKV2lGvi0OOtFetuXV3BrJMvILUoECzV/3q9hOWxVwhk9SYjuemGORUo2CSTyrdzPCUshEd8I6lisbcBPns3gk5sUqfRIm2pZDM1N8TOY2NGceh7YwpDs2iNxX/8zoZRpdBLlSaIVdsvijKJMGETJ8nfaE5Qzm2hDIt7K2EDammDG1EFRuCt/jyMvHP6ld17+681rgt0ijDERzDKXhwAQ24gSb4wEDCM7zCm/PovDjvzse8teQUM4fwB87nD4uEj90=</latexit><latexit sha1_base64="vYib4ronG48FFWnNXVHMYLu2Hrs=">AAAB7nicbVBNS8NAEJ3Ur1q/qh69LBbBU0lEUG8FL+KpgrGFNpTNdtMu3Wzi7kQooX/CiwcVr/4eb/4bt20O2vpg4PHeDDPzwlQKg6777ZRWVtfWN8qbla3tnd296v7Bg0kyzbjPEpnodkgNl0JxHwVK3k41p3EoeSscXU/91hPXRiTqHscpD2I6UCISjKKV2lGvi0OOtFetuXV3BrJMvILUoECzV/3q9hOWxVwhk9SYjuemGORUo2CSTyrdzPCUshEd8I6lisbcBPns3gk5sUqfRIm2pZDM1N8TOY2NGceh7YwpDs2iNxX/8zoZRpdBLlSaIVdsvijKJMGETJ8nfaE5Qzm2hDIt7K2EDammDG1EFRuCt/jyMvHP6ld17+681rgt0ijDERzDKXhwAQ24gSb4wEDCM7zCm/PovDjvzse8teQUM4fwB87nD4uEj90=</latexit> T = {L(x1, y1, . . . , xK, yK), (x, y) ⇠ q}<latexit sha1_base64="3SPqX5Pd//SZfknsNJ7h81baRq4=">AAACH3icbVBNS0JBFJ1nX2ZfVss2lyRQEHkvAmsRCG0iWxhoBj55zBtHHZz30cy8UB7+lDb9lTYtKqKd/6Z56qK0AzMczrn3ztzjhpxJZZoTI7Wyura+kd7MbG3v7O5l9w/uZRAJQhsk4IF4cLGknPm0oZji9CEUFHsup013cJX4zScqJAv8uhqFtO3hns+6jGClJSdbrsMl2DHcQn7oWEUYJRfYnUDJIgydaqJUC0XtalYAWzIPHsEeO9mcWTKngGVizUkOzVFzst96KIk86ivCsZQtywxVO8ZCMcLpOGNHkoaYDHCPtjT1sUdlO54uOIYTrXSgGwh9fAVT9XdHjD0pR56rKz2s+nLRS8T/vFakuuftmPlhpKhPZg91Iw4qgCQt6DBBieIjTTARTP8VSB8LTJTONKNDsBZXXiaN09JFybo7y1Vu5mmk0RE6RnlkoTKqoGtUQw1E0DN6Re/ow3gx3oxP42tWmjLmPYfoD4zJD9DSnhc=</latexit><latexit sha1_base64="3SPqX5Pd//SZfknsNJ7h81baRq4=">AAACH3icbVBNS0JBFJ1nX2ZfVss2lyRQEHkvAmsRCG0iWxhoBj55zBtHHZz30cy8UB7+lDb9lTYtKqKd/6Z56qK0AzMczrn3ztzjhpxJZZoTI7Wyura+kd7MbG3v7O5l9w/uZRAJQhsk4IF4cLGknPm0oZji9CEUFHsup013cJX4zScqJAv8uhqFtO3hns+6jGClJSdbrsMl2DHcQn7oWEUYJRfYnUDJIgydaqJUC0XtalYAWzIPHsEeO9mcWTKngGVizUkOzVFzst96KIk86ivCsZQtywxVO8ZCMcLpOGNHkoaYDHCPtjT1sUdlO54uOIYTrXSgGwh9fAVT9XdHjD0pR56rKz2s+nLRS8T/vFakuuftmPlhpKhPZg91Iw4qgCQt6DBBieIjTTARTP8VSB8LTJTONKNDsBZXXiaN09JFybo7y1Vu5mmk0RE6RnlkoTKqoGtUQw1E0DN6Re/ow3gx3oxP42tWmjLmPYfoD4zJD9DSnhc=</latexit><latexit sha1_base64="3SPqX5Pd//SZfknsNJ7h81baRq4=">AAACH3icbVBNS0JBFJ1nX2ZfVss2lyRQEHkvAmsRCG0iWxhoBj55zBtHHZz30cy8UB7+lDb9lTYtKqKd/6Z56qK0AzMczrn3ztzjhpxJZZoTI7Wyura+kd7MbG3v7O5l9w/uZRAJQhsk4IF4cLGknPm0oZji9CEUFHsup013cJX4zScqJAv8uhqFtO3hns+6jGClJSdbrsMl2DHcQn7oWEUYJRfYnUDJIgydaqJUC0XtalYAWzIPHsEeO9mcWTKngGVizUkOzVFzst96KIk86ivCsZQtywxVO8ZCMcLpOGNHkoaYDHCPtjT1sUdlO54uOIYTrXSgGwh9fAVT9XdHjD0pR56rKz2s+nLRS8T/vFakuuftmPlhpKhPZg91Iw4qgCQt6DBBieIjTTARTP8VSB8LTJTONKNDsBZXXiaN09JFybo7y1Vu5mmk0RE6RnlkoTKqoGtUQw1E0DN6Re/ow3gx3oxP42tWmjLmPYfoD4zJD9DSnhc=</latexit><latexit sha1_base64="3SPqX5Pd//SZfknsNJ7h81baRq4=">AAACH3icbVBNS0JBFJ1nX2ZfVss2lyRQEHkvAmsRCG0iWxhoBj55zBtHHZz30cy8UB7+lDb9lTYtKqKd/6Z56qK0AzMczrn3ztzjhpxJZZoTI7Wyura+kd7MbG3v7O5l9w/uZRAJQhsk4IF4cLGknPm0oZji9CEUFHsup013cJX4zScqJAv8uhqFtO3hns+6jGClJSdbrsMl2DHcQn7oWEUYJRfYnUDJIgydaqJUC0XtalYAWzIPHsEeO9mcWTKngGVizUkOzVFzst96KIk86ivCsZQtywxVO8ZCMcLpOGNHkoaYDHCPtjT1sUdlO54uOIYTrXSgGwh9fAVT9XdHjD0pR56rKz2s+nLRS8T/vFakuuftmPlhpKhPZg91Iw4qgCQt6DBBieIjTTARTP8VSB8LTJTONKNDsBZXXiaN09JFybo7y1Vu5mmk0RE6RnlkoTKqoGtUQw1E0DN6Re/ow3gx3oxP42tWmjLmPYfoD4zJD9DSnhc=</latexit> L<latexit sha1_base64="UIzmQdPm4e6dU3uEnO6VqVUUQ+8=">AAAB8XicbVBNSwMxFMzWr1q/qh69BIvgqeyKoN4KXkQ8VHBtYbuUbJptQ7PJkrwVytKf4cWDilf/jTf/jdl2D9o6EBhm3iPzJkoFN+C6305lZXVtfaO6Wdva3tndq+8fPBqVacp8qoTS3YgYJrhkPnAQrJtqRpJIsE40vi78zhPThiv5AJOUhQkZSh5zSsBKQS8hMKJE5HfTfr3hNt0Z8DLxStJAJdr9+ldvoGiWMAlUEGMCz00hzIkGTgWb1nqZYSmhYzJkgaWSJMyE+SzyFJ9YZYBjpe2TgGfq742cJMZMkshOFhHNoleI/3lBBvFlmHOZZsAknX8UZwKDwsX9eMA1oyAmlhCquc2K6YhoQsG2VLMleIsnLxP/rHnV9O7PG63bso0qOkLH6BR56AK10A1qIx9RpNAzekVvDjgvzrvzMR+tOOXOIfoD5/MH7biRPg==</latexit><latexit sha1_base64="UIzmQdPm4e6dU3uEnO6VqVUUQ+8=">AAAB8XicbVBNSwMxFMzWr1q/qh69BIvgqeyKoN4KXkQ8VHBtYbuUbJptQ7PJkrwVytKf4cWDilf/jTf/jdl2D9o6EBhm3iPzJkoFN+C6305lZXVtfaO6Wdva3tndq+8fPBqVacp8qoTS3YgYJrhkPnAQrJtqRpJIsE40vi78zhPThiv5AJOUhQkZSh5zSsBKQS8hMKJE5HfTfr3hNt0Z8DLxStJAJdr9+ldvoGiWMAlUEGMCz00hzIkGTgWb1nqZYSmhYzJkgaWSJMyE+SzyFJ9YZYBjpe2TgGfq742cJMZMkshOFhHNoleI/3lBBvFlmHOZZsAknX8UZwKDwsX9eMA1oyAmlhCquc2K6YhoQsG2VLMleIsnLxP/rHnV9O7PG63bso0qOkLH6BR56AK10A1qIx9RpNAzekVvDjgvzrvzMR+tOOXOIfoD5/MH7biRPg==</latexit><latexit sha1_base64="UIzmQdPm4e6dU3uEnO6VqVUUQ+8=">AAAB8XicbVBNSwMxFMzWr1q/qh69BIvgqeyKoN4KXkQ8VHBtYbuUbJptQ7PJkrwVytKf4cWDilf/jTf/jdl2D9o6EBhm3iPzJkoFN+C6305lZXVtfaO6Wdva3tndq+8fPBqVacp8qoTS3YgYJrhkPnAQrJtqRpJIsE40vi78zhPThiv5AJOUhQkZSh5zSsBKQS8hMKJE5HfTfr3hNt0Z8DLxStJAJdr9+ldvoGiWMAlUEGMCz00hzIkGTgWb1nqZYSmhYzJkgaWSJMyE+SzyFJ9YZYBjpe2TgGfq742cJMZMkshOFhHNoleI/3lBBvFlmHOZZsAknX8UZwKDwsX9eMA1oyAmlhCquc2K6YhoQsG2VLMleIsnLxP/rHnV9O7PG63bso0qOkLH6BR56AK10A1qIx9RpNAzekVvDjgvzrvzMR+tOOXOIfoD5/MH7biRPg==</latexit><latexit sha1_base64="UIzmQdPm4e6dU3uEnO6VqVUUQ+8=">AAAB8XicbVBNSwMxFMzWr1q/qh69BIvgqeyKoN4KXkQ8VHBtYbuUbJptQ7PJkrwVytKf4cWDilf/jTf/jdl2D9o6EBhm3iPzJkoFN+C6305lZXVtfaO6Wdva3tndq+8fPBqVacp8qoTS3YgYJrhkPnAQrJtqRpJIsE40vi78zhPThiv5AJOUhQkZSh5zSsBKQS8hMKJE5HfTfr3hNt0Z8DLxStJAJdr9+ldvoGiWMAlUEGMCz00hzIkGTgWb1nqZYSmhYzJkgaWSJMyE+SzyFJ9YZYBjpe2TgGfq742cJMZMkshOFhHNoleI/3lBBvFlmHOZZsAknX8UZwKDwsX9eMA1oyAmlhCquc2K6YhoQsG2VLMleIsnLxP/rHnV9O7PG63bso0qOkLH6BR56AK10A1qIx9RpNAzekVvDjgvzrvzMR+tOOXOIfoD5/MH7biRPg==</latexit> p(T )<latexit sha1_base64="pt9rNGiWOedYqMAbHE8mujpf7lg=">AAAB9HicbVBNS8NAFHypX7V+VT16WSxCvZREBPVW8CKeKjS20May2W7apZtN2N0oJfR/ePGg4tUf481/4ybNQVsHFoaZ93iz48ecKW3b31ZpZXVtfaO8Wdna3tndq+4f3KsokYS6JOKR7PpYUc4EdTXTnHZjSXHoc9rxJ9eZ33mkUrFItPU0pl6IR4IFjGBtpIe43g+xHhPM0/bsdFCt2Q07B1omTkFqUKA1qH71hxFJQio04VipnmPH2kux1IxwOqv0E0VjTCZ4RHuGChxS5aV56hk6McoQBZE0T2iUq783UhwqNQ19M5llVIteJv7n9RIdXHopE3GiqSDzQ0HCkY5QVgEaMkmJ5lNDMJHMZEVkjCUm2hRVMSU4i19eJu5Z46rh3J3XmrdFG2U4gmOogwMX0IQbaIELBCQ8wyu8WU/Wi/VufcxHS1axcwh/YH3+AJi2kiU=</latexit><latexit sha1_base64="pt9rNGiWOedYqMAbHE8mujpf7lg=">AAAB9HicbVBNS8NAFHypX7V+VT16WSxCvZREBPVW8CKeKjS20May2W7apZtN2N0oJfR/ePGg4tUf481/4ybNQVsHFoaZ93iz48ecKW3b31ZpZXVtfaO8Wdna3tndq+4f3KsokYS6JOKR7PpYUc4EdTXTnHZjSXHoc9rxJ9eZ33mkUrFItPU0pl6IR4IFjGBtpIe43g+xHhPM0/bsdFCt2Q07B1omTkFqUKA1qH71hxFJQio04VipnmPH2kux1IxwOqv0E0VjTCZ4RHuGChxS5aV56hk6McoQBZE0T2iUq783UhwqNQ19M5llVIteJv7n9RIdXHopE3GiqSDzQ0HCkY5QVgEaMkmJ5lNDMJHMZEVkjCUm2hRVMSU4i19eJu5Z46rh3J3XmrdFG2U4gmOogwMX0IQbaIELBCQ8wyu8WU/Wi/VufcxHS1axcwh/YH3+AJi2kiU=</latexit><latexit sha1_base64="pt9rNGiWOedYqMAbHE8mujpf7lg=">AAAB9HicbVBNS8NAFHypX7V+VT16WSxCvZREBPVW8CKeKjS20May2W7apZtN2N0oJfR/ePGg4tUf481/4ybNQVsHFoaZ93iz48ecKW3b31ZpZXVtfaO8Wdna3tndq+4f3KsokYS6JOKR7PpYUc4EdTXTnHZjSXHoc9rxJ9eZ33mkUrFItPU0pl6IR4IFjGBtpIe43g+xHhPM0/bsdFCt2Q07B1omTkFqUKA1qH71hxFJQio04VipnmPH2kux1IxwOqv0E0VjTCZ4RHuGChxS5aV56hk6McoQBZE0T2iUq783UhwqNQ19M5llVIteJv7n9RIdXHopE3GiqSDzQ0HCkY5QVgEaMkmJ5lNDMJHMZEVkjCUm2hRVMSU4i19eJu5Z46rh3J3XmrdFG2U4gmOogwMX0IQbaIELBCQ8wyu8WU/Wi/VufcxHS1axcwh/YH3+AJi2kiU=</latexit><latexit sha1_base64="pt9rNGiWOedYqMAbHE8mujpf7lg=">AAAB9HicbVBNS8NAFHypX7V+VT16WSxCvZREBPVW8CKeKjS20May2W7apZtN2N0oJfR/ePGg4tUf481/4ybNQVsHFoaZ93iz48ecKW3b31ZpZXVtfaO8Wdna3tndq+4f3KsokYS6JOKR7PpYUc4EdTXTnHZjSXHoc9rxJ9eZ33mkUrFItPU0pl6IR4IFjGBtpIe43g+xHhPM0/bsdFCt2Q07B1omTkFqUKA1qH71hxFJQio04VipnmPH2kux1IxwOqv0E0VjTCZ4RHuGChxS5aV56hk6McoQBZE0T2iUq783UhwqNQ19M5llVIteJv7n9RIdXHopE3GiqSDzQ0HCkY5QVgEaMkmJ5lNDMJHMZEVkjCUm2hRVMSU4i19eJu5Z46rh3J3XmrdFG2U4gmOogwMX0IQbaIELBCQ8wyu8WU/Wi/VufcxHS1axcwh/YH3+AJi2kiU=</latexit>
  • 20. • From line 10 in Algorithm 2, 
 
 (Recall: ) Gradient of Gradient 4. Approach
  • 21. • From line 10 in Algorithm 2, 
 
 (Recall: )
 
 
 ( is differentiable) Gradient of Gradient 4. Approach L<latexit sha1_base64="UIzmQdPm4e6dU3uEnO6VqVUUQ+8=">AAAB8XicbVBNSwMxFMzWr1q/qh69BIvgqeyKoN4KXkQ8VHBtYbuUbJptQ7PJkrwVytKf4cWDilf/jTf/jdl2D9o6EBhm3iPzJkoFN+C6305lZXVtfaO6Wdva3tndq+8fPBqVacp8qoTS3YgYJrhkPnAQrJtqRpJIsE40vi78zhPThiv5AJOUhQkZSh5zSsBKQS8hMKJE5HfTfr3hNt0Z8DLxStJAJdr9+ldvoGiWMAlUEGMCz00hzIkGTgWb1nqZYSmhYzJkgaWSJMyE+SzyFJ9YZYBjpe2TgGfq742cJMZMkshOFhHNoleI/3lBBvFlmHOZZsAknX8UZwKDwsX9eMA1oyAmlhCquc2K6YhoQsG2VLMleIsnLxP/rHnV9O7PG63bso0qOkLH6BR56AK10A1qIx9RpNAzekVvDjgvzrvzMR+tOOXOIfoD5/MH7biRPg==</latexit><latexit sha1_base64="UIzmQdPm4e6dU3uEnO6VqVUUQ+8=">AAAB8XicbVBNSwMxFMzWr1q/qh69BIvgqeyKoN4KXkQ8VHBtYbuUbJptQ7PJkrwVytKf4cWDilf/jTf/jdl2D9o6EBhm3iPzJkoFN+C6305lZXVtfaO6Wdva3tndq+8fPBqVacp8qoTS3YgYJrhkPnAQrJtqRpJIsE40vi78zhPThiv5AJOUhQkZSh5zSsBKQS8hMKJE5HfTfr3hNt0Z8DLxStJAJdr9+ldvoGiWMAlUEGMCz00hzIkGTgWb1nqZYSmhYzJkgaWSJMyE+SzyFJ9YZYBjpe2TgGfq742cJMZMkshOFhHNoleI/3lBBvFlmHOZZsAknX8UZwKDwsX9eMA1oyAmlhCquc2K6YhoQsG2VLMleIsnLxP/rHnV9O7PG63bso0qOkLH6BR56AK10A1qIx9RpNAzekVvDjgvzrvzMR+tOOXOIfoD5/MH7biRPg==</latexit><latexit sha1_base64="UIzmQdPm4e6dU3uEnO6VqVUUQ+8=">AAAB8XicbVBNSwMxFMzWr1q/qh69BIvgqeyKoN4KXkQ8VHBtYbuUbJptQ7PJkrwVytKf4cWDilf/jTf/jdl2D9o6EBhm3iPzJkoFN+C6305lZXVtfaO6Wdva3tndq+8fPBqVacp8qoTS3YgYJrhkPnAQrJtqRpJIsE40vi78zhPThiv5AJOUhQkZSh5zSsBKQS8hMKJE5HfTfr3hNt0Z8DLxStJAJdr9+ldvoGiWMAlUEGMCz00hzIkGTgWb1nqZYSmhYzJkgaWSJMyE+SzyFJ9YZYBjpe2TgGfq742cJMZMkshOFhHNoleI/3lBBvFlmHOZZsAknX8UZwKDwsX9eMA1oyAmlhCquc2K6YhoQsG2VLMleIsnLxP/rHnV9O7PG63bso0qOkLH6BR56AK10A1qIx9RpNAzekVvDjgvzrvzMR+tOOXOIfoD5/MH7biRPg==</latexit><latexit sha1_base64="UIzmQdPm4e6dU3uEnO6VqVUUQ+8=">AAAB8XicbVBNSwMxFMzWr1q/qh69BIvgqeyKoN4KXkQ8VHBtYbuUbJptQ7PJkrwVytKf4cWDilf/jTf/jdl2D9o6EBhm3iPzJkoFN+C6305lZXVtfaO6Wdva3tndq+8fPBqVacp8qoTS3YgYJrhkPnAQrJtqRpJIsE40vi78zhPThiv5AJOUhQkZSh5zSsBKQS8hMKJE5HfTfr3hNt0Z8DLxStJAJdr9+ldvoGiWMAlUEGMCz00hzIkGTgWb1nqZYSmhYzJkgaWSJMyE+SzyFJ9YZYBjpe2TgGfq742cJMZMkshOFhHNoleI/3lBBvFlmHOZZsAknX8UZwKDwsX9eMA1oyAmlhCquc2K6YhoQsG2VLMleIsnLxP/rHnV9O7PG63bso0qOkLH6BR56AK10A1qIx9RpNAzekVvDjgvzrvzMR+tOOXOIfoD5/MH7biRPg==</latexit>
  • 22. • From line 10 in Algorithm 2, 
 
 (Recall: )
 
 
 ( is differentiable) Gradient of Gradient 4. Approach L<latexit sha1_base64="UIzmQdPm4e6dU3uEnO6VqVUUQ+8=">AAAB8XicbVBNSwMxFMzWr1q/qh69BIvgqeyKoN4KXkQ8VHBtYbuUbJptQ7PJkrwVytKf4cWDilf/jTf/jdl2D9o6EBhm3iPzJkoFN+C6305lZXVtfaO6Wdva3tndq+8fPBqVacp8qoTS3YgYJrhkPnAQrJtqRpJIsE40vi78zhPThiv5AJOUhQkZSh5zSsBKQS8hMKJE5HfTfr3hNt0Z8DLxStJAJdr9+ldvoGiWMAlUEGMCz00hzIkGTgWb1nqZYSmhYzJkgaWSJMyE+SzyFJ9YZYBjpe2TgGfq742cJMZMkshOFhHNoleI/3lBBvFlmHOZZsAknX8UZwKDwsX9eMA1oyAmlhCquc2K6YhoQsG2VLMleIsnLxP/rHnV9O7PG63bso0qOkLH6BR56AK10A1qIx9RpNAzekVvDjgvzrvzMR+tOOXOIfoD5/MH7biRPg==</latexit><latexit sha1_base64="UIzmQdPm4e6dU3uEnO6VqVUUQ+8=">AAAB8XicbVBNSwMxFMzWr1q/qh69BIvgqeyKoN4KXkQ8VHBtYbuUbJptQ7PJkrwVytKf4cWDilf/jTf/jdl2D9o6EBhm3iPzJkoFN+C6305lZXVtfaO6Wdva3tndq+8fPBqVacp8qoTS3YgYJrhkPnAQrJtqRpJIsE40vi78zhPThiv5AJOUhQkZSh5zSsBKQS8hMKJE5HfTfr3hNt0Z8DLxStJAJdr9+ldvoGiWMAlUEGMCz00hzIkGTgWb1nqZYSmhYzJkgaWSJMyE+SzyFJ9YZYBjpe2TgGfq742cJMZMkshOFhHNoleI/3lBBvFlmHOZZsAknX8UZwKDwsX9eMA1oyAmlhCquc2K6YhoQsG2VLMleIsnLxP/rHnV9O7PG63bso0qOkLH6BR56AK10A1qIx9RpNAzekVvDjgvzrvzMR+tOOXOIfoD5/MH7biRPg==</latexit><latexit sha1_base64="UIzmQdPm4e6dU3uEnO6VqVUUQ+8=">AAAB8XicbVBNSwMxFMzWr1q/qh69BIvgqeyKoN4KXkQ8VHBtYbuUbJptQ7PJkrwVytKf4cWDilf/jTf/jdl2D9o6EBhm3iPzJkoFN+C6305lZXVtfaO6Wdva3tndq+8fPBqVacp8qoTS3YgYJrhkPnAQrJtqRpJIsE40vi78zhPThiv5AJOUhQkZSh5zSsBKQS8hMKJE5HfTfr3hNt0Z8DLxStJAJdr9+ldvoGiWMAlUEGMCz00hzIkGTgWb1nqZYSmhYzJkgaWSJMyE+SzyFJ9YZYBjpe2TgGfq742cJMZMkshOFhHNoleI/3lBBvFlmHOZZsAknX8UZwKDwsX9eMA1oyAmlhCquc2K6YhoQsG2VLMleIsnLxP/rHnV9O7PG63bso0qOkLH6BR56AK10A1qIx9RpNAzekVvDjgvzrvzMR+tOOXOIfoD5/MH7biRPg==</latexit><latexit sha1_base64="UIzmQdPm4e6dU3uEnO6VqVUUQ+8=">AAAB8XicbVBNSwMxFMzWr1q/qh69BIvgqeyKoN4KXkQ8VHBtYbuUbJptQ7PJkrwVytKf4cWDilf/jTf/jdl2D9o6EBhm3iPzJkoFN+C6305lZXVtfaO6Wdva3tndq+8fPBqVacp8qoTS3YgYJrhkPnAQrJtqRpJIsE40vi78zhPThiv5AJOUhQkZSh5zSsBKQS8hMKJE5HfTfr3hNt0Z8DLxStJAJdr9+ldvoGiWMAlUEGMCz00hzIkGTgWb1nqZYSmhYzJkgaWSJMyE+SzyFJ9YZYBjpe2TgGfq742cJMZMkshOFhHNoleI/3lBBvFlmHOZZsAknX8UZwKDwsX9eMA1oyAmlhCquc2K6YhoQsG2VLMleIsnLxP/rHnV9O7PG63bso0qOkLH6BR56AK10A1qIx9RpNAzekVvDjgvzrvzMR+tOOXOIfoD5/MH7biRPg==</latexit>
  • 23. • From line 10 in Algorithm 2, 
 
 (Recall: )
 
 
 ( is differentiable) Gradient of Gradient 4. Approach L<latexit sha1_base64="UIzmQdPm4e6dU3uEnO6VqVUUQ+8=">AAAB8XicbVBNSwMxFMzWr1q/qh69BIvgqeyKoN4KXkQ8VHBtYbuUbJptQ7PJkrwVytKf4cWDilf/jTf/jdl2D9o6EBhm3iPzJkoFN+C6305lZXVtfaO6Wdva3tndq+8fPBqVacp8qoTS3YgYJrhkPnAQrJtqRpJIsE40vi78zhPThiv5AJOUhQkZSh5zSsBKQS8hMKJE5HfTfr3hNt0Z8DLxStJAJdr9+ldvoGiWMAlUEGMCz00hzIkGTgWb1nqZYSmhYzJkgaWSJMyE+SzyFJ9YZYBjpe2TgGfq742cJMZMkshOFhHNoleI/3lBBvFlmHOZZsAknX8UZwKDwsX9eMA1oyAmlhCquc2K6YhoQsG2VLMleIsnLxP/rHnV9O7PG63bso0qOkLH6BR56AK10A1qIx9RpNAzekVvDjgvzrvzMR+tOOXOIfoD5/MH7biRPg==</latexit><latexit sha1_base64="UIzmQdPm4e6dU3uEnO6VqVUUQ+8=">AAAB8XicbVBNSwMxFMzWr1q/qh69BIvgqeyKoN4KXkQ8VHBtYbuUbJptQ7PJkrwVytKf4cWDilf/jTf/jdl2D9o6EBhm3iPzJkoFN+C6305lZXVtfaO6Wdva3tndq+8fPBqVacp8qoTS3YgYJrhkPnAQrJtqRpJIsE40vi78zhPThiv5AJOUhQkZSh5zSsBKQS8hMKJE5HfTfr3hNt0Z8DLxStJAJdr9+ldvoGiWMAlUEGMCz00hzIkGTgWb1nqZYSmhYzJkgaWSJMyE+SzyFJ9YZYBjpe2TgGfq742cJMZMkshOFhHNoleI/3lBBvFlmHOZZsAknX8UZwKDwsX9eMA1oyAmlhCquc2K6YhoQsG2VLMleIsnLxP/rHnV9O7PG63bso0qOkLH6BR56AK10A1qIx9RpNAzekVvDjgvzrvzMR+tOOXOIfoD5/MH7biRPg==</latexit><latexit sha1_base64="UIzmQdPm4e6dU3uEnO6VqVUUQ+8=">AAAB8XicbVBNSwMxFMzWr1q/qh69BIvgqeyKoN4KXkQ8VHBtYbuUbJptQ7PJkrwVytKf4cWDilf/jTf/jdl2D9o6EBhm3iPzJkoFN+C6305lZXVtfaO6Wdva3tndq+8fPBqVacp8qoTS3YgYJrhkPnAQrJtqRpJIsE40vi78zhPThiv5AJOUhQkZSh5zSsBKQS8hMKJE5HfTfr3hNt0Z8DLxStJAJdr9+ldvoGiWMAlUEGMCz00hzIkGTgWb1nqZYSmhYzJkgaWSJMyE+SzyFJ9YZYBjpe2TgGfq742cJMZMkshOFhHNoleI/3lBBvFlmHOZZsAknX8UZwKDwsX9eMA1oyAmlhCquc2K6YhoQsG2VLMleIsnLxP/rHnV9O7PG63bso0qOkLH6BR56AK10A1qIx9RpNAzekVvDjgvzrvzMR+tOOXOIfoD5/MH7biRPg==</latexit><latexit sha1_base64="UIzmQdPm4e6dU3uEnO6VqVUUQ+8=">AAAB8XicbVBNSwMxFMzWr1q/qh69BIvgqeyKoN4KXkQ8VHBtYbuUbJptQ7PJkrwVytKf4cWDilf/jTf/jdl2D9o6EBhm3iPzJkoFN+C6305lZXVtfaO6Wdva3tndq+8fPBqVacp8qoTS3YgYJrhkPnAQrJtqRpJIsE40vi78zhPThiv5AJOUhQkZSh5zSsBKQS8hMKJE5HfTfr3hNt0Z8DLxStJAJdr9+ldvoGiWMAlUEGMCz00hzIkGTgWb1nqZYSmhYzJkgaWSJMyE+SzyFJ9YZYBjpe2TgGfq742cJMZMkshOFhHNoleI/3lBBvFlmHOZZsAknX8UZwKDwsX9eMA1oyAmlhCquc2K6YhoQsG2VLMleIsnLxP/rHnV9O7PG63bso0qOkLH6BR56AK10A1qIx9RpNAzekVvDjgvzrvzMR+tOOXOIfoD5/MH7biRPg==</latexit>
  • 24. • From line 10 in Algorithm 2, 
 
 (Recall: )
 
 
 ( is differentiable) Gradient of Gradient 4. Approach Calculation of Hessian matrix is required.
 L<latexit sha1_base64="UIzmQdPm4e6dU3uEnO6VqVUUQ+8=">AAAB8XicbVBNSwMxFMzWr1q/qh69BIvgqeyKoN4KXkQ8VHBtYbuUbJptQ7PJkrwVytKf4cWDilf/jTf/jdl2D9o6EBhm3iPzJkoFN+C6305lZXVtfaO6Wdva3tndq+8fPBqVacp8qoTS3YgYJrhkPnAQrJtqRpJIsE40vi78zhPThiv5AJOUhQkZSh5zSsBKQS8hMKJE5HfTfr3hNt0Z8DLxStJAJdr9+ldvoGiWMAlUEGMCz00hzIkGTgWb1nqZYSmhYzJkgaWSJMyE+SzyFJ9YZYBjpe2TgGfq742cJMZMkshOFhHNoleI/3lBBvFlmHOZZsAknX8UZwKDwsX9eMA1oyAmlhCquc2K6YhoQsG2VLMleIsnLxP/rHnV9O7PG63bso0qOkLH6BR56AK10A1qIx9RpNAzekVvDjgvzrvzMR+tOOXOIfoD5/MH7biRPg==</latexit><latexit sha1_base64="UIzmQdPm4e6dU3uEnO6VqVUUQ+8=">AAAB8XicbVBNSwMxFMzWr1q/qh69BIvgqeyKoN4KXkQ8VHBtYbuUbJptQ7PJkrwVytKf4cWDilf/jTf/jdl2D9o6EBhm3iPzJkoFN+C6305lZXVtfaO6Wdva3tndq+8fPBqVacp8qoTS3YgYJrhkPnAQrJtqRpJIsE40vi78zhPThiv5AJOUhQkZSh5zSsBKQS8hMKJE5HfTfr3hNt0Z8DLxStJAJdr9+ldvoGiWMAlUEGMCz00hzIkGTgWb1nqZYSmhYzJkgaWSJMyE+SzyFJ9YZYBjpe2TgGfq742cJMZMkshOFhHNoleI/3lBBvFlmHOZZsAknX8UZwKDwsX9eMA1oyAmlhCquc2K6YhoQsG2VLMleIsnLxP/rHnV9O7PG63bso0qOkLH6BR56AK10A1qIx9RpNAzekVvDjgvzrvzMR+tOOXOIfoD5/MH7biRPg==</latexit><latexit sha1_base64="UIzmQdPm4e6dU3uEnO6VqVUUQ+8=">AAAB8XicbVBNSwMxFMzWr1q/qh69BIvgqeyKoN4KXkQ8VHBtYbuUbJptQ7PJkrwVytKf4cWDilf/jTf/jdl2D9o6EBhm3iPzJkoFN+C6305lZXVtfaO6Wdva3tndq+8fPBqVacp8qoTS3YgYJrhkPnAQrJtqRpJIsE40vi78zhPThiv5AJOUhQkZSh5zSsBKQS8hMKJE5HfTfr3hNt0Z8DLxStJAJdr9+ldvoGiWMAlUEGMCz00hzIkGTgWb1nqZYSmhYzJkgaWSJMyE+SzyFJ9YZYBjpe2TgGfq742cJMZMkshOFhHNoleI/3lBBvFlmHOZZsAknX8UZwKDwsX9eMA1oyAmlhCquc2K6YhoQsG2VLMleIsnLxP/rHnV9O7PG63bso0qOkLH6BR56AK10A1qIx9RpNAzekVvDjgvzrvzMR+tOOXOIfoD5/MH7biRPg==</latexit><latexit sha1_base64="UIzmQdPm4e6dU3uEnO6VqVUUQ+8=">AAAB8XicbVBNSwMxFMzWr1q/qh69BIvgqeyKoN4KXkQ8VHBtYbuUbJptQ7PJkrwVytKf4cWDilf/jTf/jdl2D9o6EBhm3iPzJkoFN+C6305lZXVtfaO6Wdva3tndq+8fPBqVacp8qoTS3YgYJrhkPnAQrJtqRpJIsE40vi78zhPThiv5AJOUhQkZSh5zSsBKQS8hMKJE5HfTfr3hNt0Z8DLxStJAJdr9+ldvoGiWMAlUEGMCz00hzIkGTgWb1nqZYSmhYzJkgaWSJMyE+SzyFJ9YZYBjpe2TgGfq742cJMZMkshOFhHNoleI/3lBBvFlmHOZZsAknX8UZwKDwsX9eMA1oyAmlhCquc2K6YhoQsG2VLMleIsnLxP/rHnV9O7PG63bso0qOkLH6BR56AK10A1qIx9RpNAzekVvDjgvzrvzMR+tOOXOIfoD5/MH7biRPg==</latexit>
  • 25. • From line 10 in Algorithm 2, 
 
 (Recall: )
 
 
 ( is differentiable) Gradient of Gradient 4. Approach Calculation of Hessian matrix is required.
 → MAML suggest 1st order approximation. L<latexit sha1_base64="UIzmQdPm4e6dU3uEnO6VqVUUQ+8=">AAAB8XicbVBNSwMxFMzWr1q/qh69BIvgqeyKoN4KXkQ8VHBtYbuUbJptQ7PJkrwVytKf4cWDilf/jTf/jdl2D9o6EBhm3iPzJkoFN+C6305lZXVtfaO6Wdva3tndq+8fPBqVacp8qoTS3YgYJrhkPnAQrJtqRpJIsE40vi78zhPThiv5AJOUhQkZSh5zSsBKQS8hMKJE5HfTfr3hNt0Z8DLxStJAJdr9+ldvoGiWMAlUEGMCz00hzIkGTgWb1nqZYSmhYzJkgaWSJMyE+SzyFJ9YZYBjpe2TgGfq742cJMZMkshOFhHNoleI/3lBBvFlmHOZZsAknX8UZwKDwsX9eMA1oyAmlhCquc2K6YhoQsG2VLMleIsnLxP/rHnV9O7PG63bso0qOkLH6BR56AK10A1qIx9RpNAzekVvDjgvzrvzMR+tOOXOIfoD5/MH7biRPg==</latexit><latexit sha1_base64="UIzmQdPm4e6dU3uEnO6VqVUUQ+8=">AAAB8XicbVBNSwMxFMzWr1q/qh69BIvgqeyKoN4KXkQ8VHBtYbuUbJptQ7PJkrwVytKf4cWDilf/jTf/jdl2D9o6EBhm3iPzJkoFN+C6305lZXVtfaO6Wdva3tndq+8fPBqVacp8qoTS3YgYJrhkPnAQrJtqRpJIsE40vi78zhPThiv5AJOUhQkZSh5zSsBKQS8hMKJE5HfTfr3hNt0Z8DLxStJAJdr9+ldvoGiWMAlUEGMCz00hzIkGTgWb1nqZYSmhYzJkgaWSJMyE+SzyFJ9YZYBjpe2TgGfq742cJMZMkshOFhHNoleI/3lBBvFlmHOZZsAknX8UZwKDwsX9eMA1oyAmlhCquc2K6YhoQsG2VLMleIsnLxP/rHnV9O7PG63bso0qOkLH6BR56AK10A1qIx9RpNAzekVvDjgvzrvzMR+tOOXOIfoD5/MH7biRPg==</latexit><latexit sha1_base64="UIzmQdPm4e6dU3uEnO6VqVUUQ+8=">AAAB8XicbVBNSwMxFMzWr1q/qh69BIvgqeyKoN4KXkQ8VHBtYbuUbJptQ7PJkrwVytKf4cWDilf/jTf/jdl2D9o6EBhm3iPzJkoFN+C6305lZXVtfaO6Wdva3tndq+8fPBqVacp8qoTS3YgYJrhkPnAQrJtqRpJIsE40vi78zhPThiv5AJOUhQkZSh5zSsBKQS8hMKJE5HfTfr3hNt0Z8DLxStJAJdr9+ldvoGiWMAlUEGMCz00hzIkGTgWb1nqZYSmhYzJkgaWSJMyE+SzyFJ9YZYBjpe2TgGfq742cJMZMkshOFhHNoleI/3lBBvFlmHOZZsAknX8UZwKDwsX9eMA1oyAmlhCquc2K6YhoQsG2VLMleIsnLxP/rHnV9O7PG63bso0qOkLH6BR56AK10A1qIx9RpNAzekVvDjgvzrvzMR+tOOXOIfoD5/MH7biRPg==</latexit><latexit sha1_base64="UIzmQdPm4e6dU3uEnO6VqVUUQ+8=">AAAB8XicbVBNSwMxFMzWr1q/qh69BIvgqeyKoN4KXkQ8VHBtYbuUbJptQ7PJkrwVytKf4cWDilf/jTf/jdl2D9o6EBhm3iPzJkoFN+C6305lZXVtfaO6Wdva3tndq+8fPBqVacp8qoTS3YgYJrhkPnAQrJtqRpJIsE40vi78zhPThiv5AJOUhQkZSh5zSsBKQS8hMKJE5HfTfr3hNt0Z8DLxStJAJdr9+ldvoGiWMAlUEGMCz00hzIkGTgWb1nqZYSmhYzJkgaWSJMyE+SzyFJ9YZYBjpe2TgGfq742cJMZMkshOFhHNoleI/3lBBvFlmHOZZsAknX8UZwKDwsX9eMA1oyAmlhCquc2K6YhoQsG2VLMleIsnLxP/rHnV9O7PG63bso0qOkLH6BR56AK10A1qIx9RpNAzekVvDjgvzrvzMR+tOOXOIfoD5/MH7biRPg==</latexit>
  • 26. 1st Order Approximation • Update rule of MAML: 4. Approach
  • 27. 1st Order Approximation • Update rule of MAML: 4. Approach This part needs calculating Hessian. Hessian makes MAML be slow.
  • 28. 1st Order Approximation • Update rule of MAML: • Update rule of MAML with 1st order approximation: 4. Approach
  • 29. 1st Order Approximation • Update rule of MAML: • Update rule of MAML with 1st order approximation: 4. Approach (Regard as constant)
  • 30. • From line 10 in Algorithm 2, 
 
 (Recall: )
 
 
 ( is differentiable) Recall: Gradient of Gradient 4. Approach In 1st order approximation, 
 we regard this as identity matrix I. L<latexit sha1_base64="UIzmQdPm4e6dU3uEnO6VqVUUQ+8=">AAAB8XicbVBNSwMxFMzWr1q/qh69BIvgqeyKoN4KXkQ8VHBtYbuUbJptQ7PJkrwVytKf4cWDilf/jTf/jdl2D9o6EBhm3iPzJkoFN+C6305lZXVtfaO6Wdva3tndq+8fPBqVacp8qoTS3YgYJrhkPnAQrJtqRpJIsE40vi78zhPThiv5AJOUhQkZSh5zSsBKQS8hMKJE5HfTfr3hNt0Z8DLxStJAJdr9+ldvoGiWMAlUEGMCz00hzIkGTgWb1nqZYSmhYzJkgaWSJMyE+SzyFJ9YZYBjpe2TgGfq742cJMZMkshOFhHNoleI/3lBBvFlmHOZZsAknX8UZwKDwsX9eMA1oyAmlhCquc2K6YhoQsG2VLMleIsnLxP/rHnV9O7PG63bso0qOkLH6BR56AK10A1qIx9RpNAzekVvDjgvzrvzMR+tOOXOIfoD5/MH7biRPg==</latexit><latexit sha1_base64="UIzmQdPm4e6dU3uEnO6VqVUUQ+8=">AAAB8XicbVBNSwMxFMzWr1q/qh69BIvgqeyKoN4KXkQ8VHBtYbuUbJptQ7PJkrwVytKf4cWDilf/jTf/jdl2D9o6EBhm3iPzJkoFN+C6305lZXVtfaO6Wdva3tndq+8fPBqVacp8qoTS3YgYJrhkPnAQrJtqRpJIsE40vi78zhPThiv5AJOUhQkZSh5zSsBKQS8hMKJE5HfTfr3hNt0Z8DLxStJAJdr9+ldvoGiWMAlUEGMCz00hzIkGTgWb1nqZYSmhYzJkgaWSJMyE+SzyFJ9YZYBjpe2TgGfq742cJMZMkshOFhHNoleI/3lBBvFlmHOZZsAknX8UZwKDwsX9eMA1oyAmlhCquc2K6YhoQsG2VLMleIsnLxP/rHnV9O7PG63bso0qOkLH6BR56AK10A1qIx9RpNAzekVvDjgvzrvzMR+tOOXOIfoD5/MH7biRPg==</latexit><latexit sha1_base64="UIzmQdPm4e6dU3uEnO6VqVUUQ+8=">AAAB8XicbVBNSwMxFMzWr1q/qh69BIvgqeyKoN4KXkQ8VHBtYbuUbJptQ7PJkrwVytKf4cWDilf/jTf/jdl2D9o6EBhm3iPzJkoFN+C6305lZXVtfaO6Wdva3tndq+8fPBqVacp8qoTS3YgYJrhkPnAQrJtqRpJIsE40vi78zhPThiv5AJOUhQkZSh5zSsBKQS8hMKJE5HfTfr3hNt0Z8DLxStJAJdr9+ldvoGiWMAlUEGMCz00hzIkGTgWb1nqZYSmhYzJkgaWSJMyE+SzyFJ9YZYBjpe2TgGfq742cJMZMkshOFhHNoleI/3lBBvFlmHOZZsAknX8UZwKDwsX9eMA1oyAmlhCquc2K6YhoQsG2VLMleIsnLxP/rHnV9O7PG63bso0qOkLH6BR56AK10A1qIx9RpNAzekVvDjgvzrvzMR+tOOXOIfoD5/MH7biRPg==</latexit><latexit sha1_base64="UIzmQdPm4e6dU3uEnO6VqVUUQ+8=">AAAB8XicbVBNSwMxFMzWr1q/qh69BIvgqeyKoN4KXkQ8VHBtYbuUbJptQ7PJkrwVytKf4cWDilf/jTf/jdl2D9o6EBhm3iPzJkoFN+C6305lZXVtfaO6Wdva3tndq+8fPBqVacp8qoTS3YgYJrhkPnAQrJtqRpJIsE40vi78zhPThiv5AJOUhQkZSh5zSsBKQS8hMKJE5HfTfr3hNt0Z8DLxStJAJdr9+ldvoGiWMAlUEGMCz00hzIkGTgWb1nqZYSmhYzJkgaWSJMyE+SzyFJ9YZYBjpe2TgGfq742cJMZMkshOFhHNoleI/3lBBvFlmHOZZsAknX8UZwKDwsX9eMA1oyAmlhCquc2K6YhoQsG2VLMleIsnLxP/rHnV9O7PG63bso0qOkLH6BR56AK10A1qIx9RpNAzekVvDjgvzrvzMR+tOOXOIfoD5/MH7biRPg==</latexit>
  • 31. Reinforcement Learning • Notations • Policy of an agent: such that a_t sim f_theta (x_t) • Task distribution: • For each Task: • Episode: • Transition distribution: q (K trajectories are drawn from q and f) • Episode length: H • Loss function: = expectation of sum of rewards 4. Approach f✓<latexit sha1_base64="vYib4ronG48FFWnNXVHMYLu2Hrs=">AAAB7nicbVBNS8NAEJ3Ur1q/qh69LBbBU0lEUG8FL+KpgrGFNpTNdtMu3Wzi7kQooX/CiwcVr/4eb/4bt20O2vpg4PHeDDPzwlQKg6777ZRWVtfWN8qbla3tnd296v7Bg0kyzbjPEpnodkgNl0JxHwVK3k41p3EoeSscXU/91hPXRiTqHscpD2I6UCISjKKV2lGvi0OOtFetuXV3BrJMvILUoECzV/3q9hOWxVwhk9SYjuemGORUo2CSTyrdzPCUshEd8I6lisbcBPns3gk5sUqfRIm2pZDM1N8TOY2NGceh7YwpDs2iNxX/8zoZRpdBLlSaIVdsvijKJMGETJ8nfaE5Qzm2hDIt7K2EDammDG1EFRuCt/jyMvHP6ld17+681rgt0ijDERzDKXhwAQ24gSb4wEDCM7zCm/PovDjvzse8teQUM4fwB87nD4uEj90=</latexit><latexit sha1_base64="vYib4ronG48FFWnNXVHMYLu2Hrs=">AAAB7nicbVBNS8NAEJ3Ur1q/qh69LBbBU0lEUG8FL+KpgrGFNpTNdtMu3Wzi7kQooX/CiwcVr/4eb/4bt20O2vpg4PHeDDPzwlQKg6777ZRWVtfWN8qbla3tnd296v7Bg0kyzbjPEpnodkgNl0JxHwVK3k41p3EoeSscXU/91hPXRiTqHscpD2I6UCISjKKV2lGvi0OOtFetuXV3BrJMvILUoECzV/3q9hOWxVwhk9SYjuemGORUo2CSTyrdzPCUshEd8I6lisbcBPns3gk5sUqfRIm2pZDM1N8TOY2NGceh7YwpDs2iNxX/8zoZRpdBLlSaIVdsvijKJMGETJ8nfaE5Qzm2hDIt7K2EDammDG1EFRuCt/jyMvHP6ld17+681rgt0ijDERzDKXhwAQ24gSb4wEDCM7zCm/PovDjvzse8teQUM4fwB87nD4uEj90=</latexit><latexit sha1_base64="vYib4ronG48FFWnNXVHMYLu2Hrs=">AAAB7nicbVBNS8NAEJ3Ur1q/qh69LBbBU0lEUG8FL+KpgrGFNpTNdtMu3Wzi7kQooX/CiwcVr/4eb/4bt20O2vpg4PHeDDPzwlQKg6777ZRWVtfWN8qbla3tnd296v7Bg0kyzbjPEpnodkgNl0JxHwVK3k41p3EoeSscXU/91hPXRiTqHscpD2I6UCISjKKV2lGvi0OOtFetuXV3BrJMvILUoECzV/3q9hOWxVwhk9SYjuemGORUo2CSTyrdzPCUshEd8I6lisbcBPns3gk5sUqfRIm2pZDM1N8TOY2NGceh7YwpDs2iNxX/8zoZRpdBLlSaIVdsvijKJMGETJ8nfaE5Qzm2hDIt7K2EDammDG1EFRuCt/jyMvHP6ld17+681rgt0ijDERzDKXhwAQ24gSb4wEDCM7zCm/PovDjvzse8teQUM4fwB87nD4uEj90=</latexit><latexit sha1_base64="vYib4ronG48FFWnNXVHMYLu2Hrs=">AAAB7nicbVBNS8NAEJ3Ur1q/qh69LBbBU0lEUG8FL+KpgrGFNpTNdtMu3Wzi7kQooX/CiwcVr/4eb/4bt20O2vpg4PHeDDPzwlQKg6777ZRWVtfWN8qbla3tnd296v7Bg0kyzbjPEpnodkgNl0JxHwVK3k41p3EoeSscXU/91hPXRiTqHscpD2I6UCISjKKV2lGvi0OOtFetuXV3BrJMvILUoECzV/3q9hOWxVwhk9SYjuemGORUo2CSTyrdzPCUshEd8I6lisbcBPns3gk5sUqfRIm2pZDM1N8TOY2NGceh7YwpDs2iNxX/8zoZRpdBLlSaIVdsvijKJMGETJ8nfaE5Qzm2hDIt7K2EDammDG1EFRuCt/jyMvHP6ld17+681rgt0ijDERzDKXhwAQ24gSb4wEDCM7zCm/PovDjvzse8teQUM4fwB87nD4uEj90=</latexit> p(T )<latexit sha1_base64="pt9rNGiWOedYqMAbHE8mujpf7lg=">AAAB9HicbVBNS8NAFHypX7V+VT16WSxCvZREBPVW8CKeKjS20May2W7apZtN2N0oJfR/ePGg4tUf481/4ybNQVsHFoaZ93iz48ecKW3b31ZpZXVtfaO8Wdna3tndq+4f3KsokYS6JOKR7PpYUc4EdTXTnHZjSXHoc9rxJ9eZ33mkUrFItPU0pl6IR4IFjGBtpIe43g+xHhPM0/bsdFCt2Q07B1omTkFqUKA1qH71hxFJQio04VipnmPH2kux1IxwOqv0E0VjTCZ4RHuGChxS5aV56hk6McoQBZE0T2iUq783UhwqNQ19M5llVIteJv7n9RIdXHopE3GiqSDzQ0HCkY5QVgEaMkmJ5lNDMJHMZEVkjCUm2hRVMSU4i19eJu5Z46rh3J3XmrdFG2U4gmOogwMX0IQbaIELBCQ8wyu8WU/Wi/VufcxHS1axcwh/YH3+AJi2kiU=</latexit><latexit sha1_base64="pt9rNGiWOedYqMAbHE8mujpf7lg=">AAAB9HicbVBNS8NAFHypX7V+VT16WSxCvZREBPVW8CKeKjS20May2W7apZtN2N0oJfR/ePGg4tUf481/4ybNQVsHFoaZ93iz48ecKW3b31ZpZXVtfaO8Wdna3tndq+4f3KsokYS6JOKR7PpYUc4EdTXTnHZjSXHoc9rxJ9eZ33mkUrFItPU0pl6IR4IFjGBtpIe43g+xHhPM0/bsdFCt2Q07B1omTkFqUKA1qH71hxFJQio04VipnmPH2kux1IxwOqv0E0VjTCZ4RHuGChxS5aV56hk6McoQBZE0T2iUq783UhwqNQ19M5llVIteJv7n9RIdXHopE3GiqSDzQ0HCkY5QVgEaMkmJ5lNDMJHMZEVkjCUm2hRVMSU4i19eJu5Z46rh3J3XmrdFG2U4gmOogwMX0IQbaIELBCQ8wyu8WU/Wi/VufcxHS1axcwh/YH3+AJi2kiU=</latexit><latexit sha1_base64="pt9rNGiWOedYqMAbHE8mujpf7lg=">AAAB9HicbVBNS8NAFHypX7V+VT16WSxCvZREBPVW8CKeKjS20May2W7apZtN2N0oJfR/ePGg4tUf481/4ybNQVsHFoaZ93iz48ecKW3b31ZpZXVtfaO8Wdna3tndq+4f3KsokYS6JOKR7PpYUc4EdTXTnHZjSXHoc9rxJ9eZ33mkUrFItPU0pl6IR4IFjGBtpIe43g+xHhPM0/bsdFCt2Q07B1omTkFqUKA1qH71hxFJQio04VipnmPH2kux1IxwOqv0E0VjTCZ4RHuGChxS5aV56hk6McoQBZE0T2iUq783UhwqNQ19M5llVIteJv7n9RIdXHopE3GiqSDzQ0HCkY5QVgEaMkmJ5lNDMJHMZEVkjCUm2hRVMSU4i19eJu5Z46rh3J3XmrdFG2U4gmOogwMX0IQbaIELBCQ8wyu8WU/Wi/VufcxHS1axcwh/YH3+AJi2kiU=</latexit><latexit sha1_base64="pt9rNGiWOedYqMAbHE8mujpf7lg=">AAAB9HicbVBNS8NAFHypX7V+VT16WSxCvZREBPVW8CKeKjS20May2W7apZtN2N0oJfR/ePGg4tUf481/4ybNQVsHFoaZ93iz48ecKW3b31ZpZXVtfaO8Wdna3tndq+4f3KsokYS6JOKR7PpYUc4EdTXTnHZjSXHoc9rxJ9eZ33mkUrFItPU0pl6IR4IFjGBtpIe43g+xHhPM0/bsdFCt2Q07B1omTkFqUKA1qH71hxFJQio04VipnmPH2kux1IxwOqv0E0VjTCZ4RHuGChxS5aV56hk6McoQBZE0T2iUq783UhwqNQ19M5llVIteJv7n9RIdXHopE3GiqSDzQ0HCkY5QVgEaMkmJ5lNDMJHMZEVkjCUm2hRVMSU4i19eJu5Z46rh3J3XmrdFG2U4gmOogwMX0IQbaIELBCQ8wyu8WU/Wi/VufcxHS1axcwh/YH3+AJi2kiU=</latexit> L<latexit sha1_base64="UIzmQdPm4e6dU3uEnO6VqVUUQ+8=">AAAB8XicbVBNSwMxFMzWr1q/qh69BIvgqeyKoN4KXkQ8VHBtYbuUbJptQ7PJkrwVytKf4cWDilf/jTf/jdl2D9o6EBhm3iPzJkoFN+C6305lZXVtfaO6Wdva3tndq+8fPBqVacp8qoTS3YgYJrhkPnAQrJtqRpJIsE40vi78zhPThiv5AJOUhQkZSh5zSsBKQS8hMKJE5HfTfr3hNt0Z8DLxStJAJdr9+ldvoGiWMAlUEGMCz00hzIkGTgWb1nqZYSmhYzJkgaWSJMyE+SzyFJ9YZYBjpe2TgGfq742cJMZMkshOFhHNoleI/3lBBvFlmHOZZsAknX8UZwKDwsX9eMA1oyAmlhCquc2K6YhoQsG2VLMleIsnLxP/rHnV9O7PG63bso0qOkLH6BR56AK10A1qIx9RpNAzekVvDjgvzrvzMR+tOOXOIfoD5/MH7biRPg==</latexit><latexit sha1_base64="UIzmQdPm4e6dU3uEnO6VqVUUQ+8=">AAAB8XicbVBNSwMxFMzWr1q/qh69BIvgqeyKoN4KXkQ8VHBtYbuUbJptQ7PJkrwVytKf4cWDilf/jTf/jdl2D9o6EBhm3iPzJkoFN+C6305lZXVtfaO6Wdva3tndq+8fPBqVacp8qoTS3YgYJrhkPnAQrJtqRpJIsE40vi78zhPThiv5AJOUhQkZSh5zSsBKQS8hMKJE5HfTfr3hNt0Z8DLxStJAJdr9+ldvoGiWMAlUEGMCz00hzIkGTgWb1nqZYSmhYzJkgaWSJMyE+SzyFJ9YZYBjpe2TgGfq742cJMZMkshOFhHNoleI/3lBBvFlmHOZZsAknX8UZwKDwsX9eMA1oyAmlhCquc2K6YhoQsG2VLMleIsnLxP/rHnV9O7PG63bso0qOkLH6BR56AK10A1qIx9RpNAzekVvDjgvzrvzMR+tOOXOIfoD5/MH7biRPg==</latexit><latexit sha1_base64="UIzmQdPm4e6dU3uEnO6VqVUUQ+8=">AAAB8XicbVBNSwMxFMzWr1q/qh69BIvgqeyKoN4KXkQ8VHBtYbuUbJptQ7PJkrwVytKf4cWDilf/jTf/jdl2D9o6EBhm3iPzJkoFN+C6305lZXVtfaO6Wdva3tndq+8fPBqVacp8qoTS3YgYJrhkPnAQrJtqRpJIsE40vi78zhPThiv5AJOUhQkZSh5zSsBKQS8hMKJE5HfTfr3hNt0Z8DLxStJAJdr9+ldvoGiWMAlUEGMCz00hzIkGTgWb1nqZYSmhYzJkgaWSJMyE+SzyFJ9YZYBjpe2TgGfq742cJMZMkshOFhHNoleI/3lBBvFlmHOZZsAknX8UZwKDwsX9eMA1oyAmlhCquc2K6YhoQsG2VLMleIsnLxP/rHnV9O7PG63bso0qOkLH6BR56AK10A1qIx9RpNAzekVvDjgvzrvzMR+tOOXOIfoD5/MH7biRPg==</latexit><latexit sha1_base64="UIzmQdPm4e6dU3uEnO6VqVUUQ+8=">AAAB8XicbVBNSwMxFMzWr1q/qh69BIvgqeyKoN4KXkQ8VHBtYbuUbJptQ7PJkrwVytKf4cWDilf/jTf/jdl2D9o6EBhm3iPzJkoFN+C6305lZXVtfaO6Wdva3tndq+8fPBqVacp8qoTS3YgYJrhkPnAQrJtqRpJIsE40vi78zhPThiv5AJOUhQkZSh5zSsBKQS8hMKJE5HfTfr3hNt0Z8DLxStJAJdr9+ldvoGiWMAlUEGMCz00hzIkGTgWb1nqZYSmhYzJkgaWSJMyE+SzyFJ9YZYBjpe2TgGfq742cJMZMkshOFhHNoleI/3lBBvFlmHOZZsAknX8UZwKDwsX9eMA1oyAmlhCquc2K6YhoQsG2VLMleIsnLxP/rHnV9O7PG63bso0qOkLH6BR56AK10A1qIx9RpNAzekVvDjgvzrvzMR+tOOXOIfoD5/MH7biRPg==</latexit> T = {L(e1, e2, . . . , eK), x1 ⇠ q(x), xt+1 ⇠ q(x|xt, at), H}<latexit sha1_base64="ZL5RnVdScnbZacaHZ6rzokxi/dU=">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</latexit><latexit sha1_base64="ZL5RnVdScnbZacaHZ6rzokxi/dU=">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</latexit><latexit sha1_base64="ZL5RnVdScnbZacaHZ6rzokxi/dU=">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</latexit><latexit sha1_base64="ZL5RnVdScnbZacaHZ6rzokxi/dU=">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</latexit> e = (x1, a1, . . . , xH, aH)<latexit sha1_base64="RTvl4VSKDR1Czf8GOPpJeflQwCo=">AAACDnicbVBNS8MwGE7n15xfVY9egkOZMEYrgnoQBl6GpwnWDbZS0jTbwtK0JKlslP0DL/4VLx5UvHr25r8x3XrQzQcSHp7nfd/kffyYUaks69soLC2vrK4V10sbm1vbO+bu3r2MEoGJgyMWibaPJGGUE0dRxUg7FgSFPiMtf3id+a0HIiSN+J0ax8QNUZ/THsVIackzj7t+xIKUTOAVrIw8uwpRdnWDSMkqHHmNTGiceGbZqllTwEVi56QMcjQ980uPwElIuMIMSdmxrVi5KRKKYkYmpW4iSYzwEPVJR1OOQiLddLrPBB5pJYC9SOjDFZyqvztSFEo5Dn1dGSI1kPNeJv7ndRLVu3BTyuNEEY5nD/USBlUEs3BgQAXBio01QVhQ/VeIB0ggrHSEJR2CPb/yInFOa5c1+/asXL/J0yiCA3AIKsAG56AOGqAJHIDBI3gGr+DNeDJejHfjY1ZaMPKeffAHxucP0DKZkQ==</latexit><latexit sha1_base64="RTvl4VSKDR1Czf8GOPpJeflQwCo=">AAACDnicbVBNS8MwGE7n15xfVY9egkOZMEYrgnoQBl6GpwnWDbZS0jTbwtK0JKlslP0DL/4VLx5UvHr25r8x3XrQzQcSHp7nfd/kffyYUaks69soLC2vrK4V10sbm1vbO+bu3r2MEoGJgyMWibaPJGGUE0dRxUg7FgSFPiMtf3id+a0HIiSN+J0ax8QNUZ/THsVIackzj7t+xIKUTOAVrIw8uwpRdnWDSMkqHHmNTGiceGbZqllTwEVi56QMcjQ980uPwElIuMIMSdmxrVi5KRKKYkYmpW4iSYzwEPVJR1OOQiLddLrPBB5pJYC9SOjDFZyqvztSFEo5Dn1dGSI1kPNeJv7ndRLVu3BTyuNEEY5nD/USBlUEs3BgQAXBio01QVhQ/VeIB0ggrHSEJR2CPb/yInFOa5c1+/asXL/J0yiCA3AIKsAG56AOGqAJHIDBI3gGr+DNeDJejHfjY1ZaMPKeffAHxucP0DKZkQ==</latexit><latexit sha1_base64="RTvl4VSKDR1Czf8GOPpJeflQwCo=">AAACDnicbVBNS8MwGE7n15xfVY9egkOZMEYrgnoQBl6GpwnWDbZS0jTbwtK0JKlslP0DL/4VLx5UvHr25r8x3XrQzQcSHp7nfd/kffyYUaks69soLC2vrK4V10sbm1vbO+bu3r2MEoGJgyMWibaPJGGUE0dRxUg7FgSFPiMtf3id+a0HIiSN+J0ax8QNUZ/THsVIackzj7t+xIKUTOAVrIw8uwpRdnWDSMkqHHmNTGiceGbZqllTwEVi56QMcjQ980uPwElIuMIMSdmxrVi5KRKKYkYmpW4iSYzwEPVJR1OOQiLddLrPBB5pJYC9SOjDFZyqvztSFEo5Dn1dGSI1kPNeJv7ndRLVu3BTyuNEEY5nD/USBlUEs3BgQAXBio01QVhQ/VeIB0ggrHSEJR2CPb/yInFOa5c1+/asXL/J0yiCA3AIKsAG56AOGqAJHIDBI3gGr+DNeDJejHfjY1ZaMPKeffAHxucP0DKZkQ==</latexit><latexit sha1_base64="RTvl4VSKDR1Czf8GOPpJeflQwCo=">AAACDnicbVBNS8MwGE7n15xfVY9egkOZMEYrgnoQBl6GpwnWDbZS0jTbwtK0JKlslP0DL/4VLx5UvHr25r8x3XrQzQcSHp7nfd/kffyYUaks69soLC2vrK4V10sbm1vbO+bu3r2MEoGJgyMWibaPJGGUE0dRxUg7FgSFPiMtf3id+a0HIiSN+J0ax8QNUZ/THsVIackzj7t+xIKUTOAVrIw8uwpRdnWDSMkqHHmNTGiceGbZqllTwEVi56QMcjQ980uPwElIuMIMSdmxrVi5KRKKYkYmpW4iSYzwEPVJR1OOQiLddLrPBB5pJYC9SOjDFZyqvztSFEo5Dn1dGSI1kPNeJv7ndRLVu3BTyuNEEY5nD/USBlUEs3BgQAXBio01QVhQ/VeIB0ggrHSEJR2CPb/yInFOa5c1+/asXL/J0yiCA3AIKsAG56AOGqAJHIDBI3gGr+DNeDJejHfjY1ZaMPKeffAHxucP0DKZkQ==</latexit>
  • 33. Experiment on Few-Shot Classification • Omniglot (Lake et al., 2012) • 50 different alphabets, 1623 characters. • 20 instances for each characters were drawn by 20 different people. • 1200 for training, 423 for test.
 • Mini-Imagenet (Ravi & Larochelle, 2017) • Classes for each set: train=64, validation=12, test=24.
 5. Experiment
  • 34. • MAML outperforms state-of-arts algorithms. Results of Few-Shot Classification 5. Experiment
  • 35. Effect of 1st Order Approximation • In 1st order approximation, computation is roughly 33% better. • Performances are similar. 5. Experiment
  • 36. Effect of 1st Order Approximation • In 1st order approximation, computation is roughly 33% better. • Performances are similar. 5. Experiment Performances are not that different.
  • 37. Effect of 1st Order Approximation • In 1st order approximation, computation is roughly 33% better. • Performances are similar. 5. Experiment Some performance is even better.
  • 38. Experiments on Regression • Sinusoid function: • Amplitude (A) and phase (ɸ) are varied between tasks • A in [0.1, 0.5] • ɸ in [0, π] • x in [-5.0, 5.0] • Loss function: Mean Squared Error (MSE) • Regressor: 2 hidden layers with 40 units and ReLU • Training • Use only 1 gradient step for learner • K = 5 or 10 example (5-shot learning or 10-shot learning) • Fixed step size (ɑ=0.01) for Adam optimizer. 5. Experiment
  • 39. Results of 10-Shot Learning Regression 5. Experiment [MAML] [Pretraining + Fine-tuning] The red line is ground truth. Fit this sine function with only few (10) samples.
  • 40. Results of 10-Shot Learning Regression 5. Experiment Above plots are the pre-trained function of two models. (The prediction of meta-parameter of MAML,
 The prediction of co-learned parameter of vanilla multi-task learning) [MAML] [Pretraining + Fine-tuning]
  • 41. Results of 10-Shot Learning Regression 5. Experiment After 1 gradient step update. [MAML] [Pretraining + Fine-tuning]
  • 42. Results of 10-Shot Learning Regression 5. Experiment [MAML] [Pretraining + Fine-tuning] After 10 gradient step update.
  • 43. Results of 10-Shot Learning Regression 5. Experiment * step size 0.02 In the case of MAML, they quickly adapted to new samples.
 But, adaptation failed in the case of pretraining model. [MAML] [Pretraining + Fine-tuning]
  • 44. Results of 5-Shot Learning Regression 5. Experiment * step size 0.01 [MAML] [Pretraining + Fine-tuning] In the 5-shot learning, the difference is pervasive.

  • 45. Results of 5-Shot Learning Regression 5. Experiment * step size 0.01 In the 5-shot learning, the difference is pervasive.
 Good predictions are also made for ranges not particularly seen. [MAML] [Pretraining + Fine-tuning]
  • 46. MAML Needs Only One Gradient Step 5. Experiment Vanilla pretrained model adapted slowly, but, the MAML method quickly adapted even in one gradient step.
  • 47. MAML Needs Only One Gradient Step 5. Experiment Vanilla pretrained model adapted slowly, but, the MAML method quickly adapted even in one gradient step. Meta parameter theta of MAML is more sensitive than pre-updated theta of vanilla pretrained model.
  • 48. MSE of the Meta-Parameters • The performance of the-meta parameters was not improved much in training.
 • However, the performance of 
 the single gradient updated parameters 
 started on meta-parameters
 improved as training progressed. 5. Experiment
  • 49. Experiments on Reinforcement Learning • rllab benchmark suite • Neural network policy with two hidden layers of size 100 with ReLU • Gradients updates are computed using vanilla policy gradient (REINFORCE) 
 and trust-region policy (TRPO) optimization as meta-optimizer. • Comparison • Pretraining one policy on all of the tasks and fine-tuning • Training a policy from randomly initialized weights • Oracle policy 5. Experiment
  • 50. Results on Reinforcement Learning • 2d navigation 5. Experiment
  • 51. Results on Reinforcement Learning • Locomotion • High-dimensional locomotion tasks with the MuJoCo simulator 5. Experiment
  • 52. Why 1st Order Approximation Works? 6. Discussion
  • 53. Why 1st Order Approximation Works? • [Montufar14] The objective function of ReLU network is locally linear. • So as the other piecewise linear activations (maxout, leaky relu, ...) are used. • Because, any composite function with linear or piecewise linear functions should be locally linear. • 6. Discussion magnification Solid: Shallow network with 20 hidden units Dotted: deep network (2 hidden layers) with 10 hidden units at each layer
  • 54. Why 1st Order Approximation Works? • [Montufar14] The objective function of ReLU network is locally linear. • So as the other piecewise linear activations (maxout, leaky relu, ...) are used. • Because, any composite function with linear or piecewise linear functions should be locally linear. • Hence, the Hessian matrix is almost zero.
 = Hessian matrix is meaningless. 6. Discussion magnification Solid: Shallow network with 20 hidden units Dotted: deep network (2 hidden layers) with 10 hidden units at each layer
  • 55. Why 1st Order Approximation Works? • If we change the activation function from ReLU to sigmoid? 6. Discussion
  • 56. Why 1st Order Approximation Works? • If we change the activation function from ReLU to sigmoid? 6. Discussion ReLU sigmoid w/o 1st order approx. 0.1183 0.1579 w 1st order approx. 0.1183 0.1931 ReLU sigmoid w/o 1st order approx. 61 96 w 1st order approx. 53 (13% better) 52 (46% better) [ MSE Comparison ] [ Learning Time Comparison (minutes) ]
  • 57. Why the Authors Focused on Deep Networks? • The authors said their model (MAML) is model-agnostic. • But, why they focused on deep networks? • If it is not deep networks, what kind of problem can occur? 6. Discussion
  • 58. Why the Authors Focused on Deep Networks? • Simple linear regression model (y = θx) can be used in the MAML. • Parametrized, differentiable loss function. 6. Discussion
  • 59. Why the Authors Focused on Deep Networks? • Simple linear regression model (y = θx) can be used in the MAML. • Parametrized, differentiable loss function. • Let 6. Discussion
  • 60. Why the Authors Focused on Deep Networks? • Simple linear regression model (y = θx) can be used in the MAML. • Parametrized, differentiable loss function. • Let • Then, meta parameter θ will be 1.25. 6. Discussion
  • 61. Why the Authors Focused on Deep Networks? • Let’s extend the setting. • The sample is drawn from for the task . • Then the meta parameter . 6. Discussion
  • 62. Why the Authors Focused on Deep Networks? • Let’s extend the setting. • The sample is drawn from for the task . • Then the meta parameter .
 • The MAML use fixed learning rate. • More than 2 gradient steps are needed when is far from . 6. Discussion
  • 63. Why the Authors Focused on Deep Networks? • Let’s extend the setting. • The sample is drawn from for the task . • Then the meta parameter .
 • The MAML use fixed learning rate. • More than 2 gradient steps are needed when is far from .
 • However, because the dimension of the deep model is large, 
 there may be a meta parameter
 that can quickly adapt to any task parameter . • This is why we think the authors focused on the deep models. 6. Discussion
  • 64. • However, using the deep model doesn’t eliminate the problem completely. • The authors mentioned about the generalization problem for extrapolated tasks. • The closer the task is to out-of-distribution, the worse the performance. • How can we solve this problem? Why the Authors Focused on Deep Networks? 6. Discussion
  • 65. Update Meta Parameters with N-steps • The MAML updates using only one gradient step 
 when updating the meta parameter. • The authors say that you can update meta parameters using n gradient steps.
 6. Discussion
  • 66. Update Meta Parameters with N-steps • Will it show better performance to update with N gradient steps? • If so, why the authors didn’t use this method in experiments? 6. Discussion
  • 67. Update Meta Parameters with N-steps • Will it show better performance to update with N gradient steps? • If so, why the authors didn’t use this method in experiments? • If we apply this method to MAML, the MAML model will find a sensitive meta parameter in the N-step update. 6. Discussion
  • 68. Update Meta Parameters with N-steps • Will it show better performance to update with N gradient steps? • If so, why the authors didn’t use this method in experiments? • If we apply this method to MAML, the MAML model will find a sensitive meta parameter in the N-step update. • Also, we should calculate (n + 1)-order partial derivatives. • (c.f.) Hessian matrix is 2-order partial derivatives. • The computation of (n + 1)-order partial derivatives is computationally expensive. 6. Discussion
  • 69. Update Meta Parameters with N-steps • How about using 1st order approximation with ReLU activation 
 and updating meta parameter with N-steps? 6. Discussion
  • 70. Update Meta Parameters with N-steps • N-step(s) updates with 1st order approximation. • Of course, we use ReLU for activation function. 6. Discussion 1-step 2-steps 3-steps Learning Time 53 73 92
  • 71. Update Meta Parameters with N-steps • N-step(s) updates with 1st order approximation. • Of course, we use ReLU for activation function. 6. Discussion 1-step 2-steps 3-steps Learning Time 53 73 92 the performance of the meta-parameter getting worser when n grows.
  • 72. Update Meta Parameters with N-steps • N-step(s) updates with 1st order approximation. • Of course, we use ReLU for activation function. 6. Discussion 1-step 2-steps 3-steps Learning Time 53 73 92 Final performance may be unrelated to N. (More experiments are needed for confirming this issue.)
  • 73. Unsupervised Learning • The authors said that the MAML is model agnostic. • They showed the MAML is worked on supervised learning. • How about unsupervised learning? 6. Discussion
  • 74. • The authors said that the MAML is model agnostic. • They showed the MAML is worked on supervised learning. • How about unsupervised learning? • We can use autoencoder. • Autoencoder can make unsupervised learning
 be regarded as supervised learning. Unsupervised Learning 6. Discussion
  • 75. • MAML is really slow algorithm even with 1st order approximation. • How can we make the MAML framework be faster? How to Make the MAML be faster? 6. Discussion
  • 76. • MAML is really slow algorithm even with 1st order approximation. • How can we make the MAML framework be faster? • [Zhenguo 2017] suggests 
 optimizing learning rate alpha. • Instead of putting alpha 
 as a hyperparameter, 
 train the vector alpha 
 as the parameters
 with other parameters. • It outperforms the vanilla MAML. How to Make the MAML be faster? 6. Discussion
  • 77. Application • The MAML is good for a few-shot learning problem. • Can you guess any applications of few-shot learning problem? 6. Discussion
  • 78. Application • The MAML is good for a few-shot learning problem. • Can you guess any applications of few-shot learning problem? • We think one of the most effective few-shot learning problem is 
 real estate price prediction. • There are a lot of transactions related to big apartments. • But, there are many transactions in small apartments.
 So, it is difficult to predict. • Using MAML seems to solve the problem of insufficient sample size of small apartments. • Also, other problems would be solved by MAML. 6. Discussion
  • 80. Reference • Finn, Chelsea, Pieter Abbeel, and Sergey Levine. "Model-agnostic meta-learning for fast adaptation of deep networks." arXiv preprint arXiv:1703.03400 (2017). • Montufar, Guido F., et al. "On the number of linear regions of deep neural networks." Advances in neural information processing systems. 2014. • Baldi, Pierre. "Autoencoders, unsupervised learning, and deep architectures." Proceedings of ICML workshop on unsupervised and transfer learning. 2012. • Li, Zhenguo, et al. "Meta-sgd: Learning to learn quickly for few shot learning." arXiv preprint arXiv:1707.09835 (2017)
  • 81. Reference • https://www.slideshare.net/TaesuKim3/pr12094-modelagnostic- metalearning-for-fast-adaptation-of-deep-networks • https://people.eecs.berkeley.edu/~cbfinn/_files/ nips2017_metaworkshop.pdf • https://github.com/cbfinn/maml • Lab seminar slide by Young-ki Hong.