Neural Turing Machines Tristan Deleu June 23, 2016 @tristandeleu - PowerPoint PPT Presentation

Neural Turing Machines Tristan Deleu � June 23, 2016 @tristandeleu

Deep Learning

The building blocks + Convolutional 
 Fully connected 
 Recurrent 
 Others 
 Layer Layer Layer Object Recognition Predictions Speech Recognition � � � Object Detection Language Processing � � Image Segmentation �

Examples + = Object Predictions Face detection Detection + = Speech Predictions Automatic speech recognition Recognition = + Image Predictions Image segmentation Segmentation

Examples + = Language Language Processing Processing Machine translation + = Language Predictions Sentiment analysis Processing = + + Object Language Predictions Image captioning Recognition Processing

Frameworks Lasagne Lasagne Theano Torch Tensorflow Caffe Keras Neon MXNet Chainer CNTK

Theano + Lasagne https://github.com/Lasagne/Lasagne/blob/master/examples/mnist.py

Neural Turing Machines

Recurrent Neural Network y t +1 y t − 1 y t h t +1 h t − 1 h t LSTM t +1 LSTM t − 1 LSTM t x t +1 x t − 1 x t

Memory-augmented Networks BOAT Neural Network ? Boats float on water You can’t sail against the wind Boats do not fly … • Inspired by neuroscience • Memory-augmented networks : add an external memory to neural networks to act as a knowledge base • Keep track of intermediate computations — The story to answer the question in QA problems 
 Memory Networks & Dynamic Memory Networks

Memory-augmented Networks Memory Networks Dynamic Memory Networks Neural GPU Neural Stack/Queue/DeQue Stack-augmented RNN

Turing Machine 0 1 0 1 0 0 1 1 1 0 q 0 Current state Read Operation New state Write q 0 q 1 0 0 q 0 q 0 1 0 q 1 q 0 0 1 q 1 q 1 1 0 · · ·

Neural Turing Machine 0 1 0 1 0 0 1 1 1 0 q 0 Current state Read Operation New state Write Input Output q 0 q 1 0 0 q 0 q 0 1 0 ? q 1 q 0 0 1 q 1 q 1 1 0 · · ·

Heads M t 0 1 0 1 0 0 1 1 � � � � w t Turing Machine Neural Turing Machine

Neural Turing Machine y t +1 y t − 1 y t h t +1 h t − 1 h t � � Controller FF t − 1 FF t FF t +1 � � M t − 1 M t Read heads r t +1 r t − 1 r t � � x t +1 x t − 1 x t Write heads x t +1 x t − 1 x t

Neural Turing Machine y t +1 y t − 1 y t h t +1 h t − 1 h t � � Controller LSTM t +1 LSTM t − 1 LSTM t � � M t − 1 M t Read heads r t +1 r t − 1 r t � � x t +1 x t − 1 x t Write heads x t +1 x t − 1 x t

Neural Turing Machine Input Output � NTM � � � Controller � � � � Write heads Read heads � � Memory

Open-source Library medium.com/snips-ai � github.com/snipsco/ntm-lasagne �

NTM-Lasagne

Algorithmic Tasks • Goal : Learn full algorithms only from input/output examples 
 Generate as much data as we need Input Output � ? • Strong Generalization : Generalize beyond the data the NTM has seen during training 
 Longer sequences for example ? P ( X, Y )

Copy task Inputs EOS Outputs

Training

Copy task

Copy task

Copy task Length 120

Copy task Length 150

Repeat Copy task Inputs x5 EOS Outputs

Repeat Copy task

Repeat Copy task

Associative Recall task Inputs Outputs

Associative Recall task

Associative Recall task

Priority Sort task

bAbI tasks

bAbI tasks John John Mary Mary garden garden Mary went to the garden John went to the garden Mary went back to the hallway Sandra Sandra Sandra journeyed to the bathroom John went to the hallway Mary went to the bathroom hallway hallway bathroom bathroom

bAbI tasks

Conclusion • The NTM is able to learn algorithms only from examples • It shows better generalization performances compared to other recurrent architectures 
 For example LSTMs • Fully differentiable structure 
 Drawback: generalization is still not quite perfect • New take on Artificial Intelligence 
 Trying to teach machines things they can do, the same way we would learn them • Resources Theano: http://deeplearning.net/software/theano/ • Lasagne: http://lasagne.readthedocs.io/en/latest/ • NTM-Lasagne: https://github.com/snipsco/ntm-lasagne • � June 23, 2016 @tristandeleu

Thank you � June 23, 2016 @tristandeleu