5. Deep Learning (1)
Artificial Intelligence
e.g. knowledge bases
Source: Deep Learning
(Goodfellow, Bengio, Courville)
Machine Learning
e.g. logistic regression
Representation Learning
e.g. shallow autoencoders
Deep Learning
e.g. MLPs
6. Source: Deep Learning
(Goodfellow, Bengio, Courville)
Rule-based
INPUT
OUTPUT CAT
Classic ML Deep Learning
Hand
designed
program
CAT
Hand
designed
features
Mapping
from
features
CAT
CAT
Features
Mapping
from
features
Simple
Features
+ more
abstract
features
Mapping
from
features
Representation Learning
Deep Learning (2)
11. Start of a revolution - why now?
Availability of data Processing power
(GPUs)
Improved models/
techniques
(activations, regularization,
initialization, attention)
25. WaveNet - Speech Generation
Parametric
Concatenative
WaveNet
Music #2
Music #1
Source: WaveNet: A Generative Model for Raw Audio
(Oord et al., 2016)
26. Music recommendation
Collaborative filtering vs content based
Genre
Instruments
Mood
Year
Geolocation of artist
Lyrical themes
Low level (max) - Vibrato singing
Low level (avg) - noise, distortion
Low level (avg) - Chinese pop
Similarity Playlists - Fleet Foxes
Source: Deep content-based music recommendation
(Oord et al., 2013)
29. Generative Adversarial Networks
Source: Thomas Paula (slideshare)
Goal: produce counterfeit money
that is as similar as real money
Goal: distinguish between real and
counterfeit money
Generated from Gaussian or
Normal distribution (usually)
Generated instance
Training image
generator discriminator
59. Ghosting with Deep Imitation
Learning
Source: Data-Driven Ghosting using Deep Imitation Learning
(Le et al., 2017)
Role and formation discovery
4-5-1
63. - Deep Learning is here to stay
- AI won’t kill humans (for the time being)
- AI will revolutionize labor market (already is)
- Nothing to do (1) → Learn Python
- Nothing to do (2) → Learn TensorFlow/Keras/PyTorch
- Nothing to do (3) → ML Andrew Ng (Coursera), Deep Learning (Udacity),
Stanford CS231n (Youtube), Stanford CS224d (Youtube), Oxford NLP
(Github)
- Nothing to do (4) → Practice @ Kaggle