Recurrent Neural Networks

Transcription

1 Recurrent Neural Networks Javier Béjar Deep Learning 2018/2019 Fall Master in Artificial Intelligence (FIB-UPC)

2 Introduction

3 Sequential data Many problems are described by sequences Time series Video/audio processing Natural Language Processing (translation, dialogue) Bioinformatics (DNA/protein) Process control Model the problem = Extract elements sequence dependencies DL 2018/2019 Fall - MAI - FIB 1/49

4 Long time dependences Sequences can be modeled using non sequential ML methods (e.g. Sliding windows), but All sequences must have the same length Order of the elements always matter We cannot model dependencies longer than the chosen sequence length We need methods that explicitly model time dependencies capable of: Processing arbitrary length examples Providing different mappings (one to many, many to one, many to many) DL 2018/2019 Fall - MAI - FIB 2/49

5 Input-Output mapping Usual NN One to many Many to one Many to many Recurrent NN DL 2018/2019 Fall - MAI - FIB 3/49

6 One to many - Image captioning Black and white dog jumps over bar from Deep Visual-Semantic Alignments for Generating Image Descriptions Andrej Karpathy, Li Fei-Fei DL 2018/2019 Fall - MAI - FIB 4/49

7 Many to One - Sentiment Analysis My flight was just delayed, s**t Negative Never again BA, thanks for the dreadful flight Negative We arrived on time, yeehaaa! Positive Another day, another flight Neutral Efficient, quick, delightful, always with BA Positive DL 2018/2019 Fall - MAI - FIB 5/49

8 Many to Many - Machine Translation [How, many, programmers, for, changing, a, lightbulb,?] [Wie, viele, Programmierer, zum, Wechseln, einer, Glühbirne,?] [Combien, de, programmeurs, pour, changer, une, ampoule,?] [,Cuántos, programadores, para, cambiar, una, bombilla,?] [Zenbat, bonbilla, bat, aldatzeko, programatzaileak,?] DL 2018/2019 Fall - MAI - FIB 6/49

9 Recurrent Networks

10 Recurrent Neural Networks RNN are feed forward NN with edges that span adjacent time steps (recurrent edges) At each time step nodes receive input from the current data and from the previous state This makes that input data from previous time steps can influence the output at the current time step RNN are universal function approximators (Turing Complete) DL 2018/2019 Fall - MAI - FIB 7/49

11 Recurrent Node DL 2018/2019 Fall - MAI - FIB 8/49

12 Recurrent Neural Networks Input (x) is a vector of values for time t The hidden node (h) stores the state Weights are shared through time Each step the computation uses the previous step h (t+1) = f (h (t), x t+1 ; θ) = f (f (h (t 1), x t ; θ), x t+1 ; θ) = We can think of a RNN as a deep network that stacks layers through time DL 2018/2019 Fall - MAI - FIB 9/49

13 Training RNN (unfolding) DL 2018/2019 Fall - MAI - FIB 10/49

14 Activation Functions There are different choices for the activation function to compute the hidden state, but: The hyperbolic tangent function (tanh) is a popular choice versus the usual sigmoid function Good results are also achieved using the rectified linear function (ReLU) instead DL 2018/2019 Fall - MAI - FIB 11/49

15 RNN computation a (t) = b + W h (t 1) + U x (t) (1) h (t) = tanh(a (t) ) (2) y (t) = c + V h (t) (3) b and c are bias, an additional step can be added depending on the task DL 2018/2019 Fall - MAI - FIB 12/49

16 Training RNN RNN are usually trained using backpropagation The computation is unfolded through the sequence to propagate the activations and to compute the gradient This is known as Backpropagation Through Time (BPTT) Usually the input is limited in length to reduce computational cost, this is known as Truncated BPTT Assumes that influence is limited to a time horizon DL 2018/2019 Fall - MAI - FIB 13/49

17 Recurrent NN unfolded... DL 2018/2019 Fall - MAI - FIB 14/49

18 Recurrent NN (Regresssion/Classification) Loss... DL 2018/2019 Fall - MAI - FIB 15/49

19 Recurrent NN (Sequence to Sequence) Loss Loss Loss... DL 2018/2019 Fall - MAI - FIB 16/49

20 Gradient problems Two main problems during training Exploding Gradient Vanishing Gradient Problems appear because of the sharing of weights Recurrent edge weights in combination with activation function magnify (W > 1) or shrink (W < 1) the gradient exponentially with the length of the sequence Clipping gradients and regularization are usual solutions to exploding gradient DL 2018/2019 Fall - MAI - FIB 17/49

21 Recurrent NN (Gradient problems)... DL 2018/2019 Fall - MAI - FIB 18/49

22 Recurrent NN (Gradient problems) Applying the chain rule makes that propagating the values forward and backward, the longer the sequence, the smaller the influence of the past: f t W = f t s t s t W = f t s t s t 1 s t s t 1 W = = f t s t s t s t 1 s 2 s 1 s 1 W DL 2018/2019 Fall - MAI - FIB 19/49

23 Beyond Vanilla RNN Learning long time dependencies is difficult for vanilla RNN More sophisticated recurrent architectures allow reducing gradient problems Gated RNNs introduce memory and gating mechanisms When to store information in the state How much new information changes the state DL 2018/2019 Fall - MAI - FIB 20/49

24 LSTMs

25 Long Short Term Memory units (LSTMs) LSTMs specialize on learning long time dependencies They are composed by a memory cell and control gates Gates allow regulating how much the new information changes the state and flows to the next step Forget Gate Input Gate Update Gate Output Gate DL 2018/2019 Fall - MAI - FIB 21/49

26 Long Short Term Memory units (LSTMs) An LSTM propagates a hidden state (h t ) and a cell state (c t ) The hidden state acts as a short-term memory (large updates) The cell state acts as a long-term memory (small updates) Gates control updates from layer to layer Information can flow between short-term and long-term memory DL 2018/2019 Fall - MAI - FIB 22/49

27 LSTMs DL 2018/2019 Fall - MAI - FIB 23/49

28 Role of activation functions Gates use as activation functions the sigmoid and tanh functions Their role is to perform fuzzy decisions tanh: squashes value to range [-1,1] (substract, neutral, add) sigmoid: squashes value to range [0,1] (closed, open) DL 2018/2019 Fall - MAI - FIB 24/49

29 LSTMs computations x + x tanh x f t = σ(w f [h t 1, x t ]) (Forget) tanh DL 2018/2019 Fall - MAI - FIB 25/49

30 LSTMs computations x + x tanh x f t = σ(w f [h t 1, x t ]) (Forget) i t = σ(w i [h t 1, x t ]) (Input) c t = tanh(w c [h t 1, x t ]) tanh DL 2018/2019 Fall - MAI - FIB 25/49

31 LSTMs computations x + x tanh tanh x f t = σ(w f [h t 1, x t ]) (Forget) i t = σ(w i [h t 1, x t ]) (Input) c t = tanh(w c [h t 1, x t ]) c t = f t c t 1 + i t c t (Update) DL 2018/2019 Fall - MAI - FIB 25/49

32 LSTMs computations x + x tanh tanh x f t = σ(w f [h t 1, x t ]) (Forget) i t = σ(w i [h t 1, x t ]) (Input) c t = tanh(w c [h t 1, x t ]) c t = f t c t 1 + i t c t (Update) o t = σ(w o [h t 1, x t ]) h t = o t tanh(c t ) (Output) DL 2018/2019 Fall - MAI - FIB 25/49

33 GRUs

34 Gated Recurrent Units Reduces the complexity of the LSTMs Unifies the forgetting and the update gates as a unique update gate The update gate computes how the input and previous state are combined A reset gate controls the access to the previous state near to one previous state has more effect near to zero new state (updated) has more effect DL 2018/2019 Fall - MAI - FIB 26/49

35 GRU Reset Update z t = σ(w z [h t 1, x t ]) (upd) r t = σ(w r [h t 1, x t ]) (res) h t = tanh(w h [r t h t 1, x t ]) h t = (1 z t ) h t 1 + z t h t DL 2018/2019 Fall - MAI - FIB 27/49

36 Pros/Cons

37 Vanilla RNN vs LSTMs vs GRUs Empirically Vanilla RNN underperforms on complex tasks LSTMs are widely used but GRUs are very recent (2014) There is not yet theoretical arguments in the LSTMs vs GRUs question Empirical studies do not shed light to the question (see references) DL 2018/2019 Fall - MAI - FIB 28/49

38 Vanilla RNN vs LSTMs vs GRUs Jozefowicz, R., Zaremba, W., Sutskever, I. (2015). An empirical exploration of recurrent network architectures. In Proceedings of the 32nd International Conference on Machine Learning (ICML-15) (pp ). Chung, J., Gulcehre, C., Cho, K., Bengio, Y. (2014).Empirical evaluation of gated recurrent neural networks on sequence modeling. arxiv preprint arxiv: DL 2018/2019 Fall - MAI - FIB 29/49

39 Other RNN Variations

40 Bidirectional RNNs - Back from the future In some domains it is easier to learn dependencies if information flows in both directions For instance, domains where decisions depend on the whole sequence Part of Speech tagging Machine translation Speech/handwritting recognition A RNN can be split in two, so sequences are processed in both directions DL 2018/2019 Fall - MAI - FIB 30/49

41 Regular RNNs (only forward) DL 2018/2019 Fall - MAI - FIB 31/49

42 Bidirectional RNNs (forward-backward) DL 2018/2019 Fall - MAI - FIB 32/49

43 Bidirectional RNNs - Training Both directions are independent (no interconnections) Backpropagation needs to follow the graph dependencies, not all weights can be updated at the same time Propagating Forward: First compute the RNN forward and backward pass from the inputs, then the outputs Propagating Backward: First compute the RNN forward and backward pass from the outputs, then the inputs DL 2018/2019 Fall - MAI - FIB 33/49

44 Input Sequence (N) Sequence to sequence Direct sequence association Input and output sequences have the same lengths All the outputs of the BPTT are used for the output Training is straightforward RNN RNN RNN RNN Output Sequence (N) DL 2018/2019 Fall - MAI - FIB 34/49

45 Sequence to sequence Encoder-decoder architecture Input and output sequences can have different lengths Encoder RNN summarizes input in a coding state Decoder RNN generates output from that state Different options connecting Encoder-Decoder (direct, peeking, attention) or training (teacher forcing) Inference: The sequence is generated element by element using the output of the previous step or using a beam search DL 2018/2019 Fall - MAI - FIB 35/49

46 Encoder Decoder Encoder-Decoder (Plain) Today it is raining <start> hoy está lloviendo <eos> Encoder state DL 2018/2019 Fall - MAI - FIB 36/49

47 Encoder-Decoder (Plain) - Inference When generating a sequence of discrete values, the greedy approach could not be the best policy A limited exploration of the branching alternatives is needed (beam search) The sequence with the largest joint probability is the sequence generated Hoy (0.6) Ahora (0.3) Esta (0.7) Estaba (0.2) Esta (0.8) Estaba (0.1) Lloviendo (0.9) Lloviendo (0.9) Lloviendo (0.9) Lloviendo (0.9) DL 2018/2019 Fall - MAI - FIB 37/49

48 Encoder Decoder Encoder-Decoder (Peeking) Today it is raining <start> hoy está lloviendo <eos> Encoder state DL 2018/2019 Fall - MAI - FIB 38/49

49 Encoder Decoder Encoder-Decoder (Attention) Today it is raining <start> hoy está lloviendo <eos> Encoder state Attention DL 2018/2019 Fall - MAI - FIB 39/49

50 Encoder-Decoder (Teacher Forcing) Input Sequence Today it is raining <start> hoy está lloviendo Decoder Encoder hoy está lloviendo <eos> Output Sequence Encoder state DL 2018/2019 Fall - MAI - FIB 40/49

51 Augmented neural networks RNNs are Turing complete, but it is difficult to achieve it in practice New architectures include: Data structures to store information (Read/Write Operations) RNNs control the operations Attention mechanisms Graves, Wayne, Danihelka Neural Turing Machines, ArXiv preprint arxiv: C. Olah, S. Carter, Attention and Augmented Recurrent Neural Networks, Distill, Sept 9, 2016 DL 2018/2019 Fall - MAI - FIB 41/49

52 Applications

53 Large variety of applications Time series prediction NLP Reasoning Multimodal POS tagging, Machine traslation, Question answering Caption Generation for images/video Speech recognition/generation Reinforcement learning DL 2018/2019 Fall - MAI - FIB 42/49

54 Guided Laboratory

55 Sequence prediction Sequence to value Predicting the next step of a time series Classification of time series Predicting sentiment from tweets Text generation (predicting characters) Sequence to sequence Learning to add DL 2018/2019 Fall - MAI - FIB 43/49

56 Task 1: Air Quality prediction Time series regression TS Window t Dataset: Air Quality every hour Goal: Predict wind speed next hour Training time Train/ 8784 Test/6 lag 32 Neurons /1 Layer/30 epochs 30 sec TS t+1 RNN RNN Dense DL 2018/2019 Fall - MAI - FIB 44/49

57 Task 2: Electric Devices Time series classification Power Consumption Daily power consumption of household devices (7 classes, 96 attributes) Goal: Predict household device class Training time 64 Neurons /2 Layers/30 epochs 1 min Dev Class RNN RNN Dense DL 2018/2019 Fall - MAI - FIB 45/49

58 Task 3: Sentiment analysis Tweets (neg, pos, neutral) Tweets as sequences of words Preprocess: Generate vocabulary, recode sequences Embed sequences to a more convenient space Tweet Sequence Embedding RNN RNN Training time 5000 words/ 40 dim embedding 64 Neurons /1 Layer/50 epochs 3 min Sentiment Dense DL 2018/2019 Fall - MAI - FIB 46/49

59 Task 4: Text Generation Poetry text Character prediction from text windows Preprocess: sequences of characters one hot encoding Text generation by predicting characters iteratively Training time poetry1/50 chars/3 skip 64 Neu/1 Ly/10 it/10 ep it 23 min Character Text Window One hot enc RNN RNN Dense DL 2018/2019 Fall - MAI - FIB 47/49

60 Task 5: Learning to add Predicting addition results from text Input sequence: NUMBER+NUMBER Output sequence: NUMBER Preprocess: sequences of characters one hot encoding Training time ex/3 digits 128 Neu/1 Ly/50 ep 5 min 30 sec RNN Encode + RNN Decode DL 2018/2019 Fall - MAI - FIB 48/49

61 Task 5: Learning to add One Hot enc XXX+XXX RNN (enc) RepeatVector TimeDistributed RNN (dec) Dense XXXX DL 2018/2019 Fall - MAI - FIB 49/49