PR-153: SNAIL: A Simple Neural Attentive Meta-Learner
•Download as PPTX, PDF•
1 like•512 views
- Title: SNAIL: A Simple Neural Attentive Meta-Learner - Paper: https://arxiv.org/abs/1707.03141 - Youtube: https://youtu.be/zGrwpa5-_0Y Taekmin Kim, http://github.com/tantara
Recommended
More Related Content
What's hot
What's hot (20)
Recently uploaded
Recently uploaded (20)
PR-153: SNAIL: A Simple Neural Attentive Meta-Learner
- 1. A Simple Neural AttentIve Meta- Learner PR-153 Mar 31, 2019 Taekmin Kim 1
- 2. 2
- 3. Machine Learning vs. Human ● Machine Learning ○ Try to learn data points ■ Supervised/Unsupervised Learning ■ Reinforcement Learning ● Human ○ Fast adaptation with prior knowledge ■ Few-shot learning ■ Generalization across tasks 3
- 4. ● Related Work ○ LEARNING TO REINFORCEMENT LEARN ○ RL^2 ○ MAML ○ Auto-Meta Meta-Learner? Multi-armed bandit problem https://blog.floydhub.com/meta-rl/ 4
- 5. Meta-RL ● Goal: Generalization across tasks ● Notations ○ T: Task distribution e.g., driving, multi-armed bandit problems ○ T_i: Specific task e.g., Sonata, Porsche, ... ○ x_t: state ○ a_t: action 5
- 6. RNN-based Meta-RL(Agent) ● sequence-to-sequence problem ○ refer to past experience ● Drawbacks: ○ Temporally-linear dependency https://blog.floydhub.com/meta-rl/ 6
- 7. Motivation ● Temporal(Causal) Convolution ○ depends on previous steps ● Soft Attention ○ weighted sum https://www.slideshare.net/ThomasHjeldeThoresen/temporal-convolutional-networks-dethroning-rnns-for-sequence-modelling https://medium.com/syncedreview/memory-attention-sequences-8522f531dd43 7
- 8. Temporal Convolution https://www.slideshare.net/ThomasHjeldeThoresen/temporal-convolutional-networks-dethroning-rnns-for-sequence-modelling Vanilla 1D TCs (exponential) Dilated 1D TCs Vanilla 1D Convolution Temporal Convolution(TC) 8
- 9. Attention is All you Need(2017) https://mchromiak.github.io/articles/2017/Sep/12/Transformer-Attention-is-all-you-need/#.XJ6U6-szZ0c https://medium.com/@hyponymous/paper-summary-attention-is-all-you-need-22c2c7a5e06 Q: Hidden State of Decoder K: Hidden State of Encoder V: (normalized) Weights 9 PR-049: https://www.youtube.com/watch?v=6zGgVIlStXs
- 10. Motivation ● Temporal(Causal) Convolution ○ depends on previous steps ● Soft Attention ○ weighted sum https://www.slideshare.net/ThomasHjeldeThoresen/temporal-convolutional-networks-dethroning-rnns-for-sequence-modelling https://medium.com/syncedreview/memory-attention-sequences-8522f531dd43 10
- 11. Simple Neural AttentIve Learner Building Blocks ● DenseBlock ● TCBlock ● AttentionBlock 11
- 12. Dense Block 12
- 13. TC Block 13
- 14. Attention Block Query: Hidden State of Decoder Key: Hidden State of Encoder Value: (normalized) Weights 14
- 15. 15
- 16. Simple Neural AttentIve Learner Building Blocks ● DenseBlock ● TCBlock ● AttentionBlock 16
- 17. Experiments ● Supervised Learning ○ Few-Shot Learning(Image Classification) ■ n-Way DATASET ■ m-shot ● Reinforcement Learning ○ Multi-Armed Bandits ○ Tabular MDPs ○ Continuous Control ○ Visual Navigation 17
- 19. MAML 19
- 20. MAML: Optimization-based Meta-RL 20 https://arxiv.org/abs/1703.03400 PR-094: MAML https://www.youtube.com/watch?v=fxJXXKZb-ik
- 23. Results: Tabular MDPs 23 가깝고도 먼 TRPO(이웅원 님) https://www.slideshare.net/WoongwonLee/trpo-87165690
- 24. MAML 24
- 26. Summary 26 ● SNAIL ○ Temporal Convolution ○ Soft Attention ● Meta-RL is promising ● Related Work ○ LEARNING TO REINFORCEMENT LEARN ○ RL^2 ○ MAML ○ Auto-Meta ● Materials ○ Meta-RL(Chelsea Finn): http://rail.eecs.berkeley.edu/deeprlcourse/static/slides/lec-20.pdf