Skip to content

Latest commit

 

History

History
845 lines (525 loc) · 21.1 KB

archives-lecture-rnn.md

File metadata and controls

845 lines (525 loc) · 21.1 KB

class: middle, center, title-slide

Deep Learning

Lecture: Recurrent neural networks (optional)



Prof. Gilles Louppe
g.louppe@uliege.be


Today

How to make sense of sequential data?

  • Recurrent neural networks
  • Applications
  • Beyond sequences

class: middle

Many real-world problems require to process a signal with a sequence structure.

  • Sequence classification:
    • sentiment analysis in text
    • activity/action recognition in videos
    • DNA sequence classification
  • Sequence synthesis:
    • text synthesis
    • music synthesis
    • motion synthesis
  • Sequence-to-sequence translation:
    • speech recognition
    • text translation
    • part-of-speech tagging

.footnote[Credits: Francois Fleuret, 14x050/EE559 Deep Learning, EPFL.]

???

Draw all 3 setups.


class: middle

Given a set $\mathcal{X}$, if $S(\mathcal{X})$ denotes the set of sequences of elements from $\mathcal{X}$, $$S(\mathcal{X}) = \cup_{t=1}^\infty \mathcal{X}^t,$$ then we formally define:

.grid.center[ .kol-1-2.bold[Sequence classification] .kol-1-2[$f: S(\mathcal{X}) \to \bigtriangleup^C$] ] .grid.center[ .kol-1-2.bold[Sequence synthesis] .kol-1-2[$f: \mathbb{R}^d \to S(\mathcal{X})$] ] .grid.center[ .kol-1-2.bold[Sequence-to-sequence translation] .kol-1-2[$f: S(\mathcal{X}) \to S(\mathcal{Y})$] ]


In the rest of the slides, we consider only time-indexed signal, although it generalizes to arbitrary sequences.

.footnote[Credits: Francois Fleuret, EE559 Deep Learning, EPFL.]


class: middle

Recurrent neural networks


class: middle

When the input is a sequence $\mathbf{x} \in S(\mathbb{R}^p)$ of variable length $T(\mathbf{x})$, the historical approach is to use a recurrent model which maintains a recurrent state $\mathbf{h}_t \in \mathbb{R}^q$ updated at each time step $t$.

.footnote[Credits: Francois Fleuret, EE559 Deep Learning, EPFL.]


class: middle

Formally, for $t=1, ..., T(\mathbf{x})$, $$ \mathbf{h}_t = \phi(\mathbf{x}_t, \mathbf{h}_{t-1};\theta), $$ where $\phi : \mathbb{R}^p \times \mathbb{R}^q \to \mathbb{R}^q$ and $\mathbf{h}_0 \in \mathbb{R}^q$.

Predictions can be computed at any time step $t$ from the recurrent state, $$y_t = \psi(\mathbf{h}_t;\theta),$$ with $\psi : \mathbb{R}^q \to \mathbb{R}^C$.

.footnote[Credits: Francois Fleuret, EE559 Deep Learning, EPFL.]

???

Draw the next diagrams on the black board.


class: middle

.width-93[]


count: false class: middle

.width-100[]


count: false class: middle

.width-100[]


count: false class: middle

.width-100[]


class: middle

Even though the number of steps $T$ depends on $\mathbf{x}$, this is a standard computational graph, and automatic differentiation can deal with it as usual.

In the case of recurrent neural networks, this is referred to as backpropagation through time.


class: middle

.width-100[]


Elman networks

Elman networks consist of $\phi$ and $\psi$ defined as primitive neuron units, such as logistic regression units $$ \begin{aligned} \mathbf{h}_t &= \sigma_h\left( \mathbf{W}^T_{xh} \mathbf{x}_t + \mathbf{W}^T_{hh} \mathbf{h}_{t-1} + \mathbf{b}_h \right) \\ y_t &= \sigma_y\left( \mathbf{W}_y^T \mathbf{h}_t + b_y \right) \end{aligned} $$ where $\mathbf{W}^T_{xh} \in \mathbb{R}^{p\times q}, \mathbf{W}^T_{hh} \in \mathbb{R}^{q\times q}, \mathbf{b}_{h} \in \mathbb{R}^{q}, b_{y} \in \mathbb{R}, \mathbf{h}_0=0$, and where $\sigma_h$ and $\sigma_y$ are non-linear activation functions, such as the sigmoid function, $\text{tanh}$ or $\text{ReLU}$.


class: middle

Benchmark example

Learn to recognize variable-length sequences that are palindromes. For training, we use sequences of random sizes, from $1$ to $10$.

.grid[ .kol-1-4[] .kol-1-4[ $\mathbf{x}$

$(1,2,3,2,1)$
$(2,1,2)$
$(3,4,1,2)$
$(0)$
$(1,4)$ ] .kol-1-4.center[ $y$

$1$
$1$
$0$
$1$
$0$ ] ]


class: middle

.center.width-80[]


Bidirectional RNNs

Computing the recurrent states forward in time does not make use of future input values $\mathbf{x}_{t+1:T}$, even though there are known.

  • RNNs can be made bidirectional by consuming the sequence in both directions.
  • Effectively, this amounts to run the same (single direction) RNN twice:
    • once over the original sequence $\mathbf{x}_{1:T}$,
    • once over the reversed sequence $\mathbf{x}_{T:1}$.
  • The resulting recurrent states of the bidirectional RNN is the concatenation of two resulting sequences of recurrent states.

class: middle

.center.width-80[]


Stacked RNNs

Recurrent networks can be viewed as layers producing sequences $\mathbf{h}_{1:T}^\ell$ of activations.

As for dense layers, recurrent layers can be composed in series to form a .bold[stack] of recurrent networks.


.center.width-100[]


class: middle

.center.width-80[]


Gating

When unfolded through time, the graph of computation of a recurrent network can grow very deep, and training involves dealing with vanishing gradients.

  • RNN cells should include a pass-through, or additive paths, so that the recurrent state does not go repeatedly through a squashing non-linearity.
  • This is identical to skip connections in ResNet.




.center.width-70[]

.footnote[Credits: Francois Fleuret, EE559 Deep Learning, EPFL.]


class: middle

For instance, the recurrent state update can be a per-component weighted average of its previous value $\mathbf{h}_{t-1}$ and a full update $\bar{\mathbf{h}}_t$, with the weighting $\mathbf{z}_t$ depending on the input and the recurrent state, hence acting as a forget gate.

Formally, $$ \begin{aligned} \bar{\mathbf{h}}_t &= \phi(\mathbf{x}_t, \mathbf{h}_{t-1};\theta) \\ \mathbf{z}_t &= f(\mathbf{x}_t, \mathbf{h}_{t-1};\theta) \\ \mathbf{h}_t &= \mathbf{z}_t \odot \mathbf{h}_{t-1} + (1-\mathbf{z}_t) \odot \bar{\mathbf{h}}_t. \end{aligned} $$

.footnote[Credits: Francois Fleuret, EE559 Deep Learning, EPFL.]


class: middle

.center.width-80[]


class: middle

LSTM

The long short-term memory model (LSTM; Hochreiter and Schmidhuber, 1997) is an instance of the previous gated recurrent cell, with the following changes:

  • The recurrent state is split into two parts $\mathbf{c}_t$ and $\mathbf{h}_t$, where
    • $\mathbf{c}_t$ is the cell state and
    • $\mathbf{h}_t$ is output state.
  • A forget gate $\mathbf{f}_t$ selects the cell state information to erase.
  • An input gate $\mathbf{i}_t$ selects the cell state information to update.
  • An output gate $\mathbf{o}_t$ selects the cell state information to output.



.center.width-80[]

$$ \begin{aligned} \mathbf{f}_t &= \sigma\left( \mathbf{W}_f^T [ \mathbf{h}_{t-1}, \mathbf{x}_t ] + \mathbf{b}_f \right) \end{aligned} $$




.center.width-80[]

$$ \begin{aligned} \mathbf{i}_t &= \sigma\left( \mathbf{W}_i^T [ \mathbf{h}_{t-1}, \mathbf{x}_t ] + \mathbf{b}_i \right) \\\ \bar{\mathbf{c}}_t &= \text{tanh}\left( \mathbf{W}_c^T [ \mathbf{h}_{t-1}, \mathbf{x}_t ] + \mathbf{b}_c \right) \end{aligned} $$




.center.width-80[]

$$ \begin{aligned} \mathbf{c}_t &= \mathbf{f}_t \odot \mathbf{c}_{t-1} + \mathbf{i}_t \odot \bar{\mathbf{c}}_t \end{aligned} $$




.center.width-80[]

$$ \begin{aligned} \mathbf{o}_t &= \sigma\left( \mathbf{W}_o^T [ \mathbf{h}_{t-1}, \mathbf{x}_t ] + \mathbf{b}_o \right) \\\ \mathbf{h}_t &= \mathbf{o}_t \odot \text{tanh}(\mathbf{c}_{t}) \end{aligned} $$


class: middle

.center.width-80[]


class: middle

GRU

The gated recurrent unit (GRU; Cho et al, 2014) is another gated recurrent cell. It uses two gates instead of three: an update gate $\mathbf{z}_t$ and a reset gate $\mathbf{r}_t$.

???

  • GRUs perform similarly as LSTMs for language or speech modeling sequences, but with fewer parameters.
  • However, LSTMs remain strictly stronger than GRUs.

class: middle

.center.width-70[]

$$ \begin{aligned} \mathbf{z}_t &= \sigma\left( \mathbf{W}_z^T [ \mathbf{h}_{t-1}, \mathbf{x}_t ] + \mathbf{b}_z \right) \\\ \mathbf{r}_t &= \sigma\left( \mathbf{W}_r^T [ \mathbf{h}_{t-1}, \mathbf{x}_t ] + \mathbf{b}_r \right) \\\ \bar{\mathbf{h}}_t &= \text{tanh}\left( \mathbf{W}_h^T [ \mathbf{r}_t \odot \mathbf{h}_{t-1}, \mathbf{x}_t ] + \mathbf{b}_h \right) \\\ \mathbf{h}_t &= (1-\mathbf{z}_t) \odot \mathbf{h}_{t-1} + \mathbf{z}_t \odot \bar{\mathbf{h}}_t \end{aligned} $$


class: middle

.center.width-80[]


class: middle

.center.width-80[]

.center[The models do not generalize to sequences longer than those in the training set!]


class: middle, center

(demo)


Exploding gradients

Gated units prevent gradients from vanishing, but not from exploding.


.center.width-95[![](figures/archives-lec-rnn/loss-exploding.png)]

.footnote[Credits: pat-coady.]


class: middle

.center.width-60[]

The standard strategy to solve this issue is gradient norm clipping, which rescales the norm of the gradient to a fixed threshold $\delta$ when it is above: $$\tilde{\nabla} f = \frac{\nabla f}{||\nabla f||} \min(||\nabla f||, \delta).$$

.footnote[Credits: Francois Fleuret, EE559 Deep Learning, EPFL.]


Orthogonal initialization

Let us consider a simplified RNN, with no inputs, no bias, an identity activation function $\sigma$ (as in the positive part of a ReLU) and the initial recurrent state $\mathbf{h}_0$ set to the identity matrix.

We have, $$ \begin{aligned} \mathbf{h}_t &= \sigma\left( \mathbf{W}^T_{xh} \mathbf{x}_t + \mathbf{W}^T_{hh} \mathbf{h}_{t-1} + \mathbf{b}_h \right) \\ &= \mathbf{W}^T_{hh} \mathbf{h}_{t-1} \\ &= \mathbf{W}^T \mathbf{h}_{t-1}. \end{aligned} $$

For a sequence of size $n$, it comes $$\mathbf{h}_n = \mathbf{W}(\mathbf{W}(\mathbf{W}(...(\mathbf{W}\mathbf{h}_0)...))) = \mathbf{W}^n\mathbf{h}_0 = \mathbf{W}^n I = \mathbf{W}^n.$$

Ideally, we would like $\mathbf{W}^n$ to neither vanish nor explode as $n$ increases.

???

Do on the iPad


class: middle

Fibonacci digression

The Fibonacci sequence is $$0, 1, 1, 2, 3, 5, 8, 13, 21, 34, 55, 89, 144, 233, 377, ...$$ It grows fast! But how fast?


class: middle

In matrix form, the Fibonacci sequence is equivalently expressed as

$$ \begin{pmatrix} f_{k+2} \\\ f_{k+1} \end{pmatrix} = \begin{pmatrix} 1 & 1 \\\ 1 & 0 \end{pmatrix}\begin{pmatrix} f_{k+1} \\\ f_{k} \end{pmatrix}. $$

With $\mathbf{f}_0 = \begin{pmatrix} 1 \\ 0 \end{pmatrix}$, we have $$\mathbf{f}_{k+1} = \mathbf{A} \mathbf{f}_{k} = \mathbf{A}^{k+1} \mathbf{f}_{0}.$$


class: middle

The matrix $\mathbf{A}$ can be diagonalized as $$\mathbf{A} = \mathbf{S} \Lambda \mathbf{S}^{-1},$$ where $$ \begin{aligned} \Lambda &= \begin{pmatrix} \varphi & 0 \\ 0 & -\varphi^{-1} \end{pmatrix}\\ \mathbf{S} &= \begin{pmatrix} \varphi & -\varphi^{-1} \\ 1 & 1 \end{pmatrix} \end{aligned}. $$

In particular, $$\mathbf{A}^n = \mathbf{S} \Lambda^n \mathbf{S}^{-1}.$$

Therefore, the Fibonacci sequence grows exponentially fast with the golden ratio $\varphi$.

???

$\varphi = \frac{1+\sqrt{5}}{2}$


class: middle

Theorem

Let $\rho(\mathbf{A})$ be the spectral radius of the matrix $\mathbf{A}$, defined as $$\rho(\mathbf{A}) = \max\{ |\lambda_1|,...,|\lambda_d| \}.$$

We have:

  • if $\rho(\mathbf{A}) < 1$ then $\lim_{n\to\infty} ||\mathbf{A}^n|| = \mathbf{0}$ (= vanishing activations),
  • if $\rho(\mathbf{A}) > 1$ then $\lim_{n\to\infty} ||\mathbf{A}^n|| = \infty$ (= exploding activations).

class: middle

.center[

$\rho(\mathbf{A}) < 1$, $\mathbf{A}^n$ vanish. ]

.footnote[Credits: Stephen Merety, Explaining and illustrating orthogonal initialization for recurrent neural networks, 2016.]


class: middle

.center[

$\rho(\mathbf{A}) > 1$, $\mathbf{A}^n$ explode. ]

.footnote[Credits: Stephen Merety, Explaining and illustrating orthogonal initialization for recurrent neural networks, 2016.]


class: middle

Orthogonal initialization

If $\mathbf{A}$ is orthogonal, then it is diagonalizable and all its eigenvalues are equal to $-1$ or $1$. In this case, the norm of $$\mathbf{A}^n = \mathbf{S} \Lambda^n \mathbf{S}^{-1}$$ remains bounded.

  • Therefore, initializing $\mathbf{W}$ as a random orthogonal matrix will guarantee that activations will neither vanish nor explode.
  • In practice, a random orthogonal matrix can be found through the SVD decomposition or the QR factorization of a random matrix.
  • This initialization strategy is known as orthogonal initialization.

class: middle

.center[

$\mathbf{A}$ is orthogonal. ]

.footnote[Credits: Stephen Merety, Explaining and illustrating orthogonal initialization for recurrent neural networks, 2016.]


class: middle

Exploding activations are also the reason why squashing non-linearity functions (such as $\text{tanh}$ or sigmoids) are preferred in RNNs, since they avoid recurrent states from exploding by upper bounding $||\mathbf{h}_t||$.

(At least when running the network forward.)

???

https://github.com/nyu-dl/NLP_DL_Lecture_Note/blob/master/lecture_note.pdf


class: middle

Some applications


class: middle

Sentiment analysis

.center[ .width-100[]

Document-level modeling for sentiment analysis (= text classification),
with stacked, bidirectional and gated recurrent networks. ]

.footnote[Credits: Duyu Tang et al, Document Modeling with Gated Recurrent Neural Network for Sentiment Classification, 2015.]


class: middle

Language models

Language models model language as a Markov chain, in which sentences are sequences of words $\mathbf{w}_{1:T}$ drawn repeatedly from $p(\mathbf{w}_t | \mathbf{w}_{1:t-1}).$

This is an instance of sequence synthesis, for which predictions are computed at all time steps $t$. .center[ .width-60[] ]

.footnote[Credits: Alex Graves, Generating Sequences With Recurrent Neural Networks, 2013.]


class: middle

.center[ .width-80[] ]

.footnote[Credits: Max Woolf, 2018.]


class: middle

Sequence synthesis

The same generative architecture applies to any kind of sequences. E.g., sketch-rnn-demo for sketches defined as sequences of strokes.

.center.width-80[]


exclude: true class: middle

Neural machine translation

.center[ .width-100[] ]

.footnote[Credits: Yonghui Wu et al, Google’s Neural Machine Translation System: Bridging the Gap between Human and Machine Translation, 2016.]


exclude: true class: middle

.center[ .width-100[] ]

.footnote[Credits: Yonghui Wu et al, Google’s Neural Machine Translation System: Bridging the Gap between Human and Machine Translation, 2016.]


class: middle

Text-to-speech synthesis

.width-80.center[]

.footnote[Image credits: Shen et al, 2017. arXiv:1712.05884.]


class: middle

Lip-reading in the wild

.width-80.center[]

.footnote[Image credits: Chung et al, 2016. arXiv:1611.05358.]


class: middle, black-slide

.center[

<iframe width="640" height="400" src="https://www.youtube.com/embed/5aogzAUPilE?&loop=1&start=0" frameborder="0" volume="0" allowfullscreen></iframe> ]

class: middle, black-slide

Learning to control

A recurrent network playing Mario Kart.

.center[

<iframe width="640" height="400" src="https://www.youtube.com/embed/Ipi40cb_RsI?&loop=1&start=0" frameborder="0" volume="0" allowfullscreen></iframe> ]

class: middle

Beyond sequences


class: middle

.center.circle.width-30[]

.italic[ An increasingly large number of .bold[people are defining the networks procedurally in a data-dependent way (with loops and conditionals)], allowing them to change dynamically as a function of the input data fed to them. It's really .bold[very much like a regular program, except it's parameterized]. ]

.pull-right[Yann LeCun (Director of AI Research, Facebook, 2018)]


Programs as neural nets

The topology of a recurrent network unrolled through time is not fixed, but dynamic. It depends on:

  • the input sequence and its size
  • a graph construction algorithms which consumes input tokens in sequence to add layers to the graph of computation.

class: middle

This principle generalizes to:

  • arbitrarily structured data (e.g., sequences, trees, graphs)
  • arbitrary graph of computation construction algorithms that traverses these structures (e.g., including for-loops or recursive calls).

class: middle

Neural message passing

.center[ .width-50[] .width-60[] ]

Even though the graph topology is dynamic, the unrolled computation is fully differentiable. The program is trainable.

.footnote[Credits: Henrion et al, 2017.]


class: middle

Graph neural network for object detection in point clouds

.center.width-100[]

.footnote[Credits: Shi and Rajkumar, Point-GNN, 2020.]


class: middle

Quantum chemistry with graph networks

.center.width-65[]

.footnote[Credits: Schutt et al, 2017.]

???

quantum-mechanical properties of molecular systems


class: middle

Learning to simulate physics with graph networks

.center.width-100[]

.footnote[Credits: Sanchez-Gonzalez et al, 2020.]


class: middle, black-slide

.center[

]

.footnote[Credits: Sanchez-Gonzalez et al, 2020.]


Neural computers

.center.width-55[]

.center[Any Turing machine can be simulated by a recurrent neural network
(Siegelmann and Sontag, 1995)]

???

This implies that usual programs can all be equivalently implemented as a neural network, in theory.


class: middle

.center.width-80[]

Networks can be coupled with memory storage to produce neural computers:

  • The controller processes the input sequence and interacts with the memory to generate the output.
  • The read and write operations attend to all the memory addresses.

???

  • A Turing machine is not very practical to implement useful programs.
  • However, we can simulate the general architecture of a computer in a neural network.

class: middle

.width-100[]

A differentiable neural computer being trained to store and recall dense binary numbers. Upper left: the input (red) and target (blue), as 5-bit words and a 1 bit interrupt signal. Upper right: the model's output


class: end-slide, center count: false

The end.