One shot learning - deep learning ( meta learn )
3 likes651 views
This document discusses one-shot learning and memory-augmented neural networks. It begins by explaining traditional deep learning models require large datasets for training. One-shot learning aims to give models inference ability after training on just one or few examples. Neural Turing Machines were an early approach to one-shot learning using an external memory component. Memory-Augmented Neural Networks were later developed using only content-based addressing in the external memory to enable rapid learning from small datasets.
![[N] – SHOT Learning
In deep learning models
davinnovation@gmail.com](https://image.slidesharecdn.com/oneshotlearning-170620042200/85/One-shot-learning-deep-learning-meta-learn-1-320.jpg)
![Neural Turing Machine
https://arxiv.org/abs/1410.5401
Memory
Read Heads Write Heads
Controller
External Input External Output
Neural Network
(RNN)
Write Heads
σ𝑖 𝑊𝑡(𝑖) = 1 (0 ≤ 𝑊𝑡(𝑖) ≤ 1)
0.9 0.1 0 ...
Weight vector : 𝑊𝑡0.1 1.8
2 1
0.1 2.7
erase vector : 𝑒𝑡 (0 < 𝑒𝑡 < 1)
𝑀′
𝑡(𝑖) ← 𝑀𝑡−1(𝑖)[1 − 𝑤𝑡 𝑖 𝑒𝑡]
~1
~0
~1
In my case - For compute, set `nearly’ 0 and 1 ( it’s not actually 0 || 1 )
`
Erase -> Add](https://image.slidesharecdn.com/oneshotlearning-170620042200/85/One-shot-learning-deep-learning-meta-learn-12-320.jpg)
![Active One-Shot
https://cs.stanford.edu/~woodward/papers/active_one_shot_learning_2016.pdf
RNN
𝑦𝑡, 𝑥 𝑡+1 or (0, 𝑥 𝑡+1)
[0,1] || (𝑦′ 𝑡, 0)
𝑦𝑡, 𝑥 𝑡+1
[0,1]
𝑟𝑡 = −0.05
𝑦𝑡, 𝑥 𝑡+1
(𝑦′ 𝑡, 0)
𝑟𝑡 = ቊ
+1, 𝑖𝑓 𝑦′ 𝑡 = 𝑦𝑡
−1](https://image.slidesharecdn.com/oneshotlearning-170620042200/85/One-shot-learning-deep-learning-meta-learn-33-320.jpg)
One shot learning - deep learning ( meta learn )
- 1. [N] – SHOT Learning In deep learning models davinnovation@gmail.com
- 2. Learning process Human Deep Learning
- 3. What they need… single one picture book 60,000 train data (MNIST)
- 4. In Traditional Deep Learning… http://pinktentacle.com/tag/waseda-university/
- 5. One Shot Learning Give Inference ability to Learning model
- 6. One Shot Learning Train One(Few) Data, Works Awesome
- 7. One Shot Learning – Many approaches Transfer Learning Domain Adaptation Imitation Learning …..
- 8. Started from Meta Learning Meta means something that is "self-referencing". ‘Knowing about knowing’ Today…
- 9. Neural Turing Machine https://arxiv.org/abs/1410.5401 Memory Read Heads Write Heads Controller External Input External Output Turing Machine Basic Structure
- 10. Neural Turing Machine https://arxiv.org/abs/1410.5401 Memory Read Heads Write Heads Controller External Input External Output Neural Network (RNN) N x M memory Block 𝑀𝑡 : t is time N M
- 11. Neural Turing Machine https://arxiv.org/abs/1410.5401 Memory Write Heads Controller External Input External Output Read Heads σ𝑖 𝑤𝑡(𝑖) = 1 (0 ≤ 𝑤𝑡(𝑖) ≤ 1)0.9 0.1 0 ... Weight vector : 𝑤𝑡 1 2 2 1 1 3 1.1 1.9 1.2 𝑟𝑡 ← 𝑖 𝑤𝑡(𝑖)𝑀𝑡(𝑖) Read vector : 𝑟𝑡 i = 0 i = 1 … i = 0 i = 1 …
- 12. Neural Turing Machine https://arxiv.org/abs/1410.5401 Memory Read Heads Write Heads Controller External Input External Output Neural Network (RNN) Write Heads σ𝑖 𝑊𝑡(𝑖) = 1 (0 ≤ 𝑊𝑡(𝑖) ≤ 1) 0.9 0.1 0 ... Weight vector : 𝑊𝑡0.1 1.8 2 1 0.1 2.7 erase vector : 𝑒𝑡 (0 < 𝑒𝑡 < 1) 𝑀′ 𝑡(𝑖) ← 𝑀𝑡−1(𝑖)[1 − 𝑤𝑡 𝑖 𝑒𝑡] ~1 ~0 ~1 In my case - For compute, set `nearly’ 0 and 1 ( it’s not actually 0 || 1 ) ` Erase -> Add
- 13. Neural Turing Machine https://arxiv.org/abs/1410.5401 Memory Read Heads Write Heads Controller External Input External Output Neural Network (RNN) Write Heads σ𝑖 𝑊𝑡(𝑖) = 1 (0 ≤ 𝑊𝑡(𝑖) ≤ 1) 0.9 0.1 0 ... Weight vector : 𝑊𝑡 1 1.9 2.9 1.1 0.1 2.7 add vector : 𝑎 𝑡 (0 < 𝑎 𝑡 < 1) 𝑀𝑡 𝑖 ← 𝑀′ 𝑡 𝑖 + 𝑤𝑡 𝑖 𝑎 𝑡 ~1 ~1 ~0 In my case - For compute, set `nearly’ 0 and 1 ( it’s not actually 0 || 1 ) Erase -> Add
- 14. Neural Turing Machine https://arxiv.org/abs/1410.5401 Memory Read Heads Write Heads Controller External Input External Output Neural Network (RNN) Addressing Mechanism How to calculate ‘Weight Vector’ 𝑊𝑡 Content-based addressing > - Based on Similarity of ‘Current value’ and ‘predicted by controller value’ location-based addressing > Like 𝑓 𝑥, 𝑦 = 𝑥 x 𝑦 : variable 𝑥 , 𝑦 store them in different addresses, retrieve them and perform a multiplication algorithm => - Based on addressed by location Use both mechanisms concurrently
- 15. Neural Turing Machine https://arxiv.org/abs/1410.5401 Memory Read Heads Write Heads Controller External Input External Output Neural Network (RNN) Addressing Mechanism How to calculate ‘Weight Vector’ 𝑊𝑡 Use both mechanisms concurrently 𝑤𝑡−1 𝑀𝑡 𝑘 𝑡 β 𝑡 𝑔𝑡 𝑠𝑡 γ 𝑡 Content Addressing Interpolation Convolutional Shift Sharpening Previous State Controller Outputs 𝑤𝑡 𝑐 𝑤𝑡 𝑔 𝑤𝑡 𝑤′ 𝑡 𝑘 𝑡 : key vector β 𝑡 : key strength 𝑔𝑡 : interpolation gate (0, 1) 𝑠𝑡 : shift weighting ( only integer ) γ 𝑡 : sharping weighting Content-based addressing Location-based addressing
- 16. Neural Turing Machine https://arxiv.org/abs/1410.5401 Memory Read Heads Write Heads Controller External Input External Output Neural Network (RNN) Addressing Mechanism How to calculate ‘Weight Vector’ 𝑊𝑡 Use both mechanisms concurrently 𝑤𝑡−1 𝑀𝑡 𝑘 𝑡 β 𝑡 𝑔𝑡 𝑠𝑡 γ 𝑡 Content Addressing Interpolation Convolutional Shift Sharpening Previous State Controller Outputs 𝑤𝑡 𝑐 𝑤𝑡 𝑔 𝑤𝑡 𝑤′ 𝑡 𝑘 𝑡 : key vector β 𝑡 : key strength 𝑔𝑡 : interpolation gate 𝑠𝑡 : shift weighting γ 𝑡 : sharping weighting Content Addressing 𝑤𝑡 𝑐 𝑘 𝑡 β 𝑡 𝑀𝑡 𝑤𝑡 𝑐 𝑖 ← exp(β 𝑡 𝐾 𝑘 𝑡,𝑀𝑡 𝑖 ) σ 𝑗 exp(β 𝑡 𝐾 𝑘 𝑡,𝑀𝑡 𝑗 ) 𝐾 u, v = 𝑢 ∙ 𝑣 | 𝑢 | ∙ | 𝑣 | ∶ 𝑐𝑜𝑠𝑖𝑛𝑒 𝑠𝑖𝑚𝑖𝑙𝑎𝑟𝑖𝑡𝑦 1 2 2 1 1 3 3 2 1 𝑘 𝑡 𝑀𝑡 β 𝑡 = 0 0.5 0.5 .61 .39 β 𝑡 = 5 .98 .02 β 𝑡 = 50 𝑤𝑡 𝑐
- 17. Neural Turing Machine https://arxiv.org/abs/1410.5401 Memory Read Heads Write Heads Controller External Input External Output Neural Network (RNN) Addressing Mechanism How to calculate ‘Weight Vector’ 𝑊𝑡 Use both mechanisms concurrently 𝑤𝑡−1 𝑀𝑡 𝑘 𝑡 β 𝑡 𝑔𝑡 𝑠𝑡 γ 𝑡 Content Addressing Interpolation Convolutional Shift Sharpening Previous State Controller Outputs 𝑤𝑡 𝑐 𝑤𝑡 𝑔 𝑤𝑡 𝑤′ 𝑡 𝑘 𝑡 : key vector β 𝑡 : key strength 𝑔𝑡 : interpolation gate 𝑠𝑡 : shift weighting γ 𝑡 : sharping weighting Interpolation𝑤𝑡 𝑐 𝑤𝑡 𝑔 𝑤𝑡−1 𝑔𝑡 𝑤𝑡 𝑔 ← 𝑔𝑡 𝑤𝑡 𝑐 + (1 − 𝑔𝑡) 𝑤𝑡−1 .61 .39 𝑤𝑡 𝑐 .01 .90 𝑤𝑡−1 𝑔𝑡 = 0 .01 .90 𝑔𝑡 = 0.5 .36 .64 𝑤𝑡 𝑔
- 18. Neural Turing Machine https://arxiv.org/abs/1410.5401 Memory Read Heads Write Heads Controller External Input External Output Neural Network (RNN) Addressing Mechanism How to calculate ‘Weight Vector’ 𝑊𝑡 Use both mechanisms concurrently 𝑤𝑡−1 𝑀𝑡 𝑘 𝑡 β 𝑡 𝑔𝑡 𝑠𝑡 γ 𝑡 Content Addressing Interpolation Convolutional Shift Sharpening Previous State Controller Outputs 𝑤𝑡 𝑐 𝑤𝑡 𝑔 𝑤𝑡 𝑤′ 𝑡 𝑘 𝑡 : key vector β 𝑡 : key strength 𝑔𝑡 : interpolation gate 𝑠𝑡 : shift weighting γ 𝑡 : sharping weighting Convolutional Shift 𝑤𝑡 𝑔 𝑤′ 𝑡 𝑠𝑡 𝑤′ 𝑡(𝑖) ← σ 𝑗=0 𝑁−1 𝑤𝑡 𝑔 𝑗 𝑠𝑡(𝑖 − 𝑗) // circular convolution .01 .90 𝑤𝑡 𝑔 0 1 𝑠𝑡 0.9 𝑤′ 𝑡(0) .01 .90 𝑤𝑡 𝑔 0 1 𝑠𝑡 .01 𝑤′ 𝑡(1) .01 .90 𝑤′ 𝑡 0 1 𝑠𝑡 1 0 𝑠𝑡 0.5 0.5 𝑠𝑡 Right shift Left shift Spread shift
- 19. Neural Turing Machine https://arxiv.org/abs/1410.5401 Memory Read Heads Write Heads Controller External Input External Output Neural Network (RNN) Addressing Mechanism How to calculate ‘Weight Vector’ 𝑊𝑡 Use both mechanisms concurrently 𝑤𝑡−1 𝑀𝑡 𝑘 𝑡 β 𝑡 𝑔𝑡 𝑠𝑡 γ 𝑡 Content Addressing Interpolation Convolutional Shift Sharpening Previous State Controller Outputs 𝑤𝑡 𝑐 𝑤𝑡 𝑔 𝑤𝑡 𝑤′ 𝑡 𝑘 𝑡 : key vector β 𝑡 : key strength 𝑔𝑡 : interpolation gate (0, 1) 𝑠𝑡 : shift weighting ( only integer ) γ 𝑡 : sharping weighting Sharpening 𝑤𝑡 𝑤′ 𝑡 γ 𝑡 𝑤𝑡 𝑖 ← 𝑤′ 𝑡(𝑖)γ 𝑡 σ 𝑗 𝑤′ 𝑡(𝑗)γ 𝑡 .01 .90 𝑤′ 𝑡 γ 𝑡 = 0 0.5 0.5 .01 .90 .0001 .9998 γ 𝑡 = 1 γ 𝑡 = 2
- 20. Neural Turing Machine https://arxiv.org/abs/1410.5401 Memory Read Heads Write Heads Controller External Input External Output Neural Network (RNN) Neural Network (RNN+LSTM) Controller
- 21. Experiment – Copy train time
- 22. Experiment – Copy result NTM LSTM Input : Random binary vector ( 8bit ) NTM > LSTM Quick & Clear
- 23. Memory-Augmented Neural Networks https://arxiv.org/pdf/1605.06065.pdf NTM is ‘Meta learning’ => Use Memory-Augment, make rapid learn model.
- 24. Memory-Augmented Neural Networks https://arxiv.org/pdf/1605.06065.pdf Memory Read Heads Write Heads Controller External Input External OutputAddressing Mechanism How to calculate ‘Weight Vector’ 𝑊𝑡 Content-based addressing > - Based on Similarity of ‘Current value’ and ‘predicted by controller value’ location-based addressing > Like 𝑓 𝑥, 𝑦 = 𝑥 x 𝑦 : variable 𝑥 , 𝑦 store them in different addresses, retrieve them and perform a multiplication algorithm => - Based on addressed by location Use both mechanisms concurrently => Doesn’t use location-based addressing => Only Use Content-based addressing Memory-Augmented Neural Networks (MANN)
- 25. Memory Read Heads Write Heads Controller External Input External Output Neural Network (RNN) Addressing Mechanism How to calculate ‘Weight Vector’ 𝑊𝑡 𝑤𝑡−1 𝑀𝑡 𝑘 𝑡 β 𝑡 𝑔𝑡 𝑠𝑡 γ 𝑡 Content Addressing Interpolation Convolutional Shift Sharpening Previous State Controller Outputs 𝑤𝑡 𝑐 𝑤𝑡 𝑔 𝑤𝑡 𝑤′ 𝑡 𝑘 𝑡 : key vector β 𝑡 : key strength 𝑔𝑡 : interpolation gate (0, 1) 𝑠𝑡 : shift weighting ( only integer ) γ 𝑡 : sharping weighting Content-based addressing Location-based addressing Memory-Augmented Neural Networks https://arxiv.org/pdf/1605.06065.pdf MANN
- 26. Memory Read Heads Write Heads Controller External Input External Output Neural Network (RNN) Addressing Mechanism How to calculate ‘Weight Vector’ 𝑊𝑡 𝑤𝑡−1 𝑀𝑡 𝑘 𝑡 β 𝑡 𝑔𝑡 𝑠𝑡 γ 𝑡 Content Addressing Interpolation Convolutional Shift Sharpening Previous State Controller Outputs 𝑤𝑡 𝑟 𝑤𝑡 𝑔 𝑤𝑡 𝑤′ 𝑡 𝑘 𝑡 : key vector β 𝑡 : key strength 𝑔𝑡 : interpolation gate (0, 1) 𝑠𝑡 : shift weighting ( only integer ) γ 𝑡 : sharping weighting Content-based addressing Location-based addressing Memory-Augmented Neural Networks https://arxiv.org/pdf/1605.06065.pdf MANN
- 27. Memory Read Heads Write Heads Controller External Input External Output Neural Network (RNN) Addressing Mechanism How to calculate ‘Weight Vector’ 𝑊𝑡 Use both mechanisms concurrently 𝑤𝑡−1 𝑀𝑡 𝑘 𝑡 β 𝑡 𝑔𝑡 𝑠𝑡 γ 𝑡 Content Addressing Interpolation Convolutional Shift Sharpening Previous State Controller Outputs 𝑤𝑡 𝑐 𝑤𝑡 𝑔 𝑤𝑡 𝑤′ 𝑡 𝑘 𝑡 : key vector β 𝑡 : key strength 𝑔𝑡 : interpolation gate 𝑠𝑡 : shift weighting γ 𝑡 : sharping weighting Content Addressing 𝑤𝑡 𝑟 𝑘 𝑡 𝑀𝑡 𝑤𝑡 𝑟 𝑖 ← exp(𝐾 𝑘 𝑡,𝑀𝑡 𝑖 ) σ 𝑗 exp(𝐾 𝑘 𝑡,𝑀𝑡 𝑗 ) 𝐾 u, v = 𝑢 ∙ 𝑣 | 𝑢 | ∙ | 𝑣 | ∶ 𝑐𝑜𝑠𝑖𝑛𝑒 𝑠𝑖𝑚𝑖𝑙𝑎𝑟𝑖𝑡𝑦 1 2 2 1 1 3 3 2 1 𝑘 𝑡 𝑀𝑡 .53 .47 𝑤𝑡 𝑟 Memory-Augmented Neural Networks https://arxiv.org/pdf/1605.06065.pdf MANN
- 28. Memory Read Heads Write Heads Controller External Input External Output Neural Network (RNN) Write Heads 0.9 0.1 0 ... 0.1 1.8 2 1 0.1 2.7 𝑤𝑡 𝑢 ← γ 𝑤𝑡−1 𝑢 + 𝑤𝑡 𝑟 + 𝑤𝑡 𝑤 ~1 ~0 ~1 In my case - For compute, set `nearly’ 0 and 1 ( it’s not actually 0 || 1 ) ` Memory-Augmented Neural Networks https://arxiv.org/pdf/1605.06065.pdf Least Recently Used Access ( LRUA ) Compute from MANN 𝑤𝑡 𝑤 ← σ α 𝑤𝑡−1 𝑟 + (1 − σ α ) 𝑤𝑡−1 𝑙𝑢 Decay Parameter 𝑤𝑡 𝑙𝑢 (𝑖) = ቊ 0, 𝑖𝑓 𝑤𝑡 𝑢 𝑖 𝑖𝑠 𝑏𝑖𝑔 𝑒𝑛𝑜𝑢𝑔ℎ 1, 𝑒𝑙𝑠𝑒 𝑀𝑡 𝑖 ← 𝑀𝑡−1 𝑖 + 𝑤𝑡 𝑤 𝑖 𝑘 𝑡 Least Used Memory Sigmoid Function Scalar gate parameter 𝑤𝑡 𝑤 𝑘 𝑡 : write vector
- 29. Experiment – Data One Episode : Input : (𝑥0, 𝑛𝑢𝑙𝑙), 𝑥1, 𝑦0 , 𝑥2, 𝑦1 , … (𝑥 𝑇, 𝑦 𝑇−1) Output : (𝑦′0), 𝑦′1 , 𝑦′2 , … (𝑦′ 𝑇) 𝑝 𝑦𝑡 𝑥𝑡; 𝐷1:𝑡−1; 𝜃) (𝑥0, 𝑛𝑢𝑙𝑙 ) (𝑦′0) RNN 𝑥1, 𝑦0 (𝑦′1)
- 30. Experiment – Data Omniglot Dataset : 1600 > classes 1200 class train, 423 class test ( downscale to 20x20 ) + plus rotate augmentation
- 31. Experiment – Classification Result Trained with one-hot vector representations With Five randomly chosen labels, train 100,000 episode ( each episode are ‘new class’ ) Instance : class emerge count…?
- 32. Active One-Shot https://cs.stanford.edu/~woodward/papers/active_one_shot_learning_2016.pdf MANN + ‘decision predict or pass’ Just Like Quiz Show!
- 33. Active One-Shot https://cs.stanford.edu/~woodward/papers/active_one_shot_learning_2016.pdf RNN 𝑦𝑡, 𝑥 𝑡+1 or (0, 𝑥 𝑡+1) [0,1] || (𝑦′ 𝑡, 0) 𝑦𝑡, 𝑥 𝑡+1 [0,1] 𝑟𝑡 = −0.05 𝑦𝑡, 𝑥 𝑡+1 (𝑦′ 𝑡, 0) 𝑟𝑡 = ቊ +1, 𝑖𝑓 𝑦′ 𝑡 = 𝑦𝑡 −1
- 34. Results https://cs.stanford.edu/~woodward/papers/active_one_shot_learning_2016.pdf
- 36. Reference • https://tensorflow.blog/tag/one-shot-learning/ • http://www.modulabs.co.kr/DeepLAB_library/11115 • https://www.youtube.com/watch?v=CzQSQ_0Z-QU • https://www.slideshare.net/JisungDavidKim/oneshot- learning • https://norman3.github.io/papers/docs/neural_turing_ machine.html • https://www.slideshare.net/webaquebec/guillaume- chevalier-deep-learning-avec-tensor-flow • https://www.slideshare.net/ssuser06e0c5/metalearnin g-with-memory-augmented-neural-networks