8 Sequence Models

8.3.1 Graphical Representation of RNN

Copy link

The following figure exhibits the difference between the classical classic feed forward neural network architecture (left figure) and the Recurrent Neural Network:

• Feed forward neural network also called vanilla neural network presents one input and provides one output without taking into account the sequence of the data.

• Recurrent Neural Network introduce an internal loop wich allows information to be passed from one step of the network to the next. It apply a recurrence relation at every time step to process a sequence of data.

8.3.2 Unfold Representation

Copy link

The unfold representation of the previous RNN is presented in the following figure which make more clear how the RNN can handle sequence data using the recurrence relation

h<t>⏟cell state=fW(h<t−1>⏟old state,x<t>⏟input vector)$$\underset{\text{cell state}}{\underset{}{h^{<t>}}}=f_W(\underset{\text{old state}}{\underset{}{h^{<t-1>}}},\underset{\text{input vector}}{\underset{}{x^{<t>}}})$$ where fW$f_W$ is a function parameterized by the weight W$W$. It is important to note that at every time step t$t$ the same function fW$f_W$ is used and same set of weight parameters.

Figure 8.8: Multilayer Layer RNN source link

8.3.3 Weight Matrices parameters in RNN Sequence Modeling Tasks

Copy link

Let introduce the weight parameters in the previous figure and precise some notation.

Below are the details associated with each parameter in the network.

• x<1>,…,x<t−1>,x<t>,x<t+1>,...x<m>$x^{<1>},\dots ,x^{<t-1>},x^{<t>},x^{<t+1>},...x^{<m>}$: the input data. In natural langage processing (NLP), for example, we can consider that the sequence input is a sentence of m$m$ words x<1>,…,x<t−1>,x<t>,x<t+1>,...x<m>$x^{<1>},\dots ,x^{<t-1>},x^{<t>},x^{<t+1>},...x^{<m>}$.

• x<t>∈R|V|$x^{<t>}\in R^{|V|}$: input vector at time t$t$. For example each word x<t>$x^{<t>}$ in the sentence will be input as 1-hot vector of size |V|$|V|$ (where V$V$ is the vocabulary and |V|$|V|$ denotes the size).

• h<t>=σ(Whhh<t−1>+Whxx<t>)$h^{<t>}=\sigma (W_{hh}{h}^{<t-1>}+W_{hx}{x}^{<t>})$: the relationship to compute the hidden layer output features at each time-step t$t$.

• σ(⋅)$\sigma (\cdot )$: the non-linearity function (sigmoid here)

• Whx∈RDh×|V|$W_{hx}\in R^{D_h\times |V|}$: weight matrix used to condition the input vector, x<t>$x^{<t>}$. The number of neurons in each layer is Dh$D_h$

• Whh∈RDh×Dh$W_{hh}\in R^{D_h\times D_h}$: weights matrix used to condition the output of the previous time-step, h<t−1>$h^{<t-1>}$

• h<t−1>∈RDh$h^{<t-1>}\in R^{D_h}$: output of the non-linear function at the previous time-step, t−1$t-1$. h<0>∈RDh$h^{<0>}\in R^{D_h}$ is an initialization vector for the hidden layer at time-step t=0$t=0$.

• Each desired output y<t>$y^{<t>}$ is a 1-hot vector of size |V|$|V|$ with 1 at index of x<t+1>$x^{<t+1>}$ in V$V$ and 0 elsewhere.

• Wyh∈∈R|V|×Dh$W_{yh}\in \in R^{|V|\times D_h}$: weights matrix used to compute the prediction at each time t$t$. For example, ˆy<t>=softmax(Wyhht)${\hat{y}}^{<t>}=softmax\left(W_{yh}{h}_t\right)$ and ˆy∈R|V|$\hat{y}\in R^{|V|}$.

8.3.4 Different types of Sequence Modeling Tasks

Copy link

Different tasks can be achieved using RNN:

• many-to-one: Sentiment Classification, action prediction (sequence of video -> action class)

• one-to-many: Image captioning (image -> sequence of words)

• many-to-many: Video Captioning (Sequance of video frames -> Caption)

• many-to-many: Video Classification on frame level

8.3.5 How to train a RNN ?

Copy link

RNN a clearly some nice features:

• can process any length input

• In theory for each time t$t$ can use information from many steps back

• Same weights applied on every timestep

However RNN could be very slow to train and in practice it is difficult to access information from many steps back.

Let consider training a RNN for language model example:

• We first have to get access to a big corpus of text which is a sequence of words x<1>,…,x<t−1>,x<t>,x<t+1>,...x<T>$x^{<1>},\dots ,x^{<t-1>},x^{<t>},x^{<t+1>},...x^{<T>}$

• Forward pass into the RNN model and compute the output distribution ˆy<t>${\hat{y}}^{<t>}$ for every time t$t$ which represent the predict probability distribution of every word , given words so far

• Compute the Loss fuction on each step t$t$. In this case the cross-entropy between predicted probability distribution ˆy<t>${\hat{y}}^{<t>}$ and the true next word y<t>$y^{<t>}$ (one-hot for ˆx<t+1>${\hat{x}}^{<t+1>}$):

L<t>(θ)=CE(y<t>,ˆy<t>)=−|V|∑j=1y<t>j×log(ˆy<t>j)$$L^{<t>}\left(\theta \right)=CE(y^{<t>},{\hat{y}}^{<t>})=-\sum _{j=1}^{|V|}{y}_j^{<t>}\times log\left({\hat{y}}_j^{<t>}\right)$$

• Average this to get overall loss for entire training set:

L=1TT∑t=1L<t>(θ)=−1TT∑t=1|V|∑j=1y<t>j×log(ˆy<t>j)$$L=\frac{1}{T}\sum _{t=1}^T{L}^{<t>}\left(\theta \right)=-\frac{1}{T}\sum _{t=1}^T\sum _{j=1}^{|V|}{y}_j^{<t>}\times log\left({\hat{y}}_j^{<t>}\right)$$

Computiong the loss and the gradients accross entire corpus is too expensive. Then , in practice we use a batch of sentences to compute the loss and Stochastic Gradient Descent is exploited to compute the gradients for samll chunk of data, and then update.

8.3.6 Backpropagation through time

Copy link

Let explicit the derivative of L$L$ w.r.t. the repeated weight matrix Whh$W_{hh}$.

The backpropagation over time as two summations:

• First summation over L$L$

∂L∂Whh=T∑j=1∂L<j>∂Whh$$\frac{\partial L}{\partial W_{hh}}=\sum _{j=1}^T\frac{\partial L^{<j>}}{\partial W_{hh}}$$ - Second summations showing that each L<t>$L^{<t>}$ depends on the weight matrices before it: ∂L<t>∂Whh=t∑i=1∂L<t>∂Whh$$\frac{\partial L^{<t>}}{\partial W_{hh}}=\sum _{i=1}^t\frac{\partial L^{<t>}}{\partial W_{hh}}$$

Using the multivariate chain-rule:

∂L<t>∂Whh=t∑k=1∂L<t>∂ˆy<t>∂ˆy<t>∂h<t>∂h<t>∂h<k>∂h<k>∂Whh$$\frac{\partial L^{<t>}}{\partial W_{hh}}=\sum _{k=1}^t\frac{\partial L^{<t>}}{\partial {\hat{y}}^{<t>}}\frac{\partial {\hat{y}}^{<t>}}{\partial h^{<t>}}\frac{\partial h^{<t>}}{\partial h^{<k>}}\frac{\partial h^{<k>}}{\partial W_{hh}}$$

Now using the chain rule differentiation over all hidden layers whitin the time interval [k,t]:

∂h<t>∂h<k>=t∏j=k+1∂h<j>∂h<j−1>=t∏j=k+1WThh×diag[f′(Whhh<j−1>+Whxx<j>)]

where each ∂h<j>∂h<j−1> is the Jacobian matrix for h:

∂h<j>∂h<j−1>=[∂h<j>∂h<j−1>1,…,∂h<j>∂h<j−1>Dh]=[∂h<j>1∂h<j−1>1...∂h<j>1∂h<j−1>Dh.........∂hj,Dh∂hj−1,1...∂hj,Dh∂hj−1,Dn]

Repeated matrix multiplication is required to get ∂h<t>∂h<k>. It leads to a vanishing or exploding gradients issues.

8.3.7 Vanishing and Exploding gradient

Copy link

Remind us that the RNN is based on this recursive relation h<t>=σ(Whhh<t−1>+Whxx<t>).

Let consider the identidy function σ(x)=x, to help to understand why we have an issue with the computation of the gradient.

∂h<t>∂h<t−1>=diag(σ′(Whhh<t−1>+Whxx<t>))Whh=IWhh=Whh

Let consider the gradient of the loss L<i> on time i, with respect to the hidden state h<j> on some previous time j and denote ℓ=i−j. Then,

∂L<i>∂h<j>=∂L<i>∂h<i>∏j<t≤i∂h<t>∂h<t−1>=∂L<i>∂h<i>∏j<t≤iWhh=∂L<i>∂h<i>∏j<t≤iWℓhh

Then if Whh is “small”, then this term gets exponentially problematic as ℓ becomes large.

Indeed, if the weight matrix Whh is diagonalizable: Whh=Q−1×Δ×Q, where Q is composed of the eigenvectors and Δ is a diagonal matrix with the eigenvalues on the diagonal. Computing the power of Whh is then given by:

Wℓhh=Q−1×Δℓ×Q.

Thus eigenvalues lower than 1 will lead to vanishing gradient while eigenvalues greater than 1 will lead to exploding gradient.

8.3.8 Remarks

Copy link

Note that gradient signal from faraway is lost because it’s much smaller than gradient signal from close-by. Thus weights are updated only with respect to near effects, not long-term effects. For example in langage modelling, the contribution of faraway words to predicting the next word at time-step t diminishes when the gradient vanishes early on.

The gradient can be viewed as a measure of the effect of the past on the future. Thus if gradient is small, the model cannot learn this dependency and the model unable to predict similar long-distance dependencies at test time.

Exploding gradient is also a big issue for updating the weight and can cause too big step during the stochastic gradient descent. Moreover, once the gradient value grows extremely large, it causes an overflow (i.e. NaN).

A solution to solve the problem of exploding gradients has been first introduced by Thomas Mikolov who proposed the Gradient clipping. The principle is to scale down the gradient before applying an update when the norm of the gradient is greater than some threshold.

8.4.1 Main ideas

Copy link

At each time t, LSTM provides a hidden state (h(t)) and a cell state (c(t)) which are both vectors of length n. The cell has the ability to stores long-tern information. Further, the LSTM model can erase, write and read information from the cell.

Gates are defined to get the ability to selection which information to either erased, written or read. The gates are also vectors of length n. At each step t the gates can be open (1), closed (0) or somewhere in-between. Note that the gates are dynamic.

8.4.2 Graphical Representation of LSTM

Copy link

Figure 8.12: Understand LSTM respresentation source link

From a sequence of inputs x(t), LSTM computes a sequence of hidden states h(t), and cell state c(t):

• Forget gate: controls what is kept versus forgotten, from previous cell state

f(t)=σ(Wfh(t−1)+Ufx(t)+bf)

• Input gate: controls what parts of the new cell content are written to cell

i(t)=σ(Wih(t−1)+Uix(t)+bi)

• Output gate: controls what parts of cell are output to hidden state

o(t)=σ(Wo(rt∘h(t−1))+Uox(t)+bo)

• New cell content: this is the new content to be written to the cell

˜c(t)=tanh(Wch(t−1)+Uc(xt+bc)

• cell state: erase (“forget”) some content from last cell state, and write (“input”) some new cell content

c(t)=f(t)∘c(t−1)+i(t)∘˜c(t)

• Hidden state: read (“output”) some content from the cell

h(t)=o(t)∘tanh c(t)

8.4.3 Remarks

Copy link

• LSTM have been largely exploited for handwritting recognition, speech recognition, machine translation, parsing, image captioning.

• The architecture of LSTM model is especially designed to preserve information over many timesteps. Indeed, LST maintains a separate cell state from what of outputted.

• LSTM uses gates to control the flow of information

• Backpropagating from c(t) to c(t−1) is only element-wise multiplication by the f gate, and there is no matrix multiplication by W. The f gate is different at every time step, ranged between 0 and 1 due to sigmoid property, thus it overcome the issue of multiplying the same thing over and over again.

• Backpropagating from h(t) to h(t−1) is going through only one single tanh nonlinearity rather than tanh for every single step.

• The Backpropagation through time with uninterrupted gradient flow helps for the vanishing gradient issue even if LSTM does not guarantee vanishing/exploding gradient issues.

8.5.1 To go further

Copy link

A interesting paper to explore comparison of LSTM and GRU models (Jozefowicz et al 2015)

8.6.1 Cho’s model

Copy link

Principle. In this model the encoder and decoder are both GRUs. The final state of the encoder is used as the summary c and then this summary is accessed by all steps in the decoder

8.6.2 Sutskever’s model

Copy link

Principle. In this model the encoder and decoder are multilayered LSTMs. The final state of the encoder becomes the initial state of the decoder. So the source sentence has to be reversed.

8.6.3 Improvement with Attention layer

Copy link

Limitation of these models are exposed in the paper “On the Properties of Neural Machine Translation: Encoder-Decoder Approaches”. They showed that the performance of the encoder-decoder network degrades rapidly as the length of the input sentence increases.

The main drawback of the previous models is that the encoded vector need to capture the entire phrase (sentence) and might skipped many important details. Moreover the information needs to “flow” through many RNN steps which is quite difficult for long sentence.

Bahdanau et al. (2015) have proposed to include attention layer which consist to include attention mechanisms to give more importance to some of the input words compared to others while translating the sentence.

A survey of the different implementations of attention model is presented by Galassi et al. (2019)

8.7.1 Further reading for MT

Copy link

• Understand attention concept

• A comprehensive illustration of the transformer

• Chen et al. (2018) who compared several types of recent MT models

• A comprehensive PyTorch-based seq2seq tutorial by Joost Bastings:The Annotated Encoder–Decoder with Attention