Recurrent networks and beyond by Tomas Mikolov

Recurrent Networks and Beyond
Tomas Mikolov, Facebook
Neu-IR Workshop, Pisa, Italy 2016

Goals of this talk
• Explain recent success of recurrent networks
• Understand better the concept of (longer) short term memory
• Explore limitations of recurrent networks
• Discuss what needs to be done to build machines that can
understand language
Tomas Mikolov, Facebook, 2016

Brief History of Recurrent Nets – 80’s & 90’s
• Recurrent network architectures were very popular in the 80’s and
early 90’s (Elman, Jordan, Mozer, Hopfield, Parallel Distributed
Processing group, …)
• The main idea is very attractive: to re-use parameters and
computation (usually over time)

Simple RNN Architecture
• Input layer, hidden layer with recurrent
connections, and the output layer
• In theory, the hidden layer can learn
to represent unlimited memory
• Also called Elman network
(Finding structure in time, Elman 1990)

Brief History of Recurrent Nets – 90’s - 2010
• After the initial excitement, recurrent nets vanished from the
mainstream research
• Despite being theoretically powerful models, RNNs were mostly
considered as unstable to be trained
• Some success was achieved at IDSIA with the Long Short Term
Memory RNN architecture, but this model was too complex for others
to reproduce easily

Brief History of Recurrent Nets – 2010 - today
• In 2010, it was shown that RNNs can significantly improve state-of-
the-art in language modeling, machine translation, data compression
and speech recognition (including strong commercial speech
recognizer from IBM)
• RNNLM toolkit was published to allow researchers to reproduce the
results and extend the techniques
• The key novel trick in RNNLM was trivial: to clip gradients to prevent
instability of training

• 21% - 24% reduction of WER on Wall Street Journal setup

• Improvement from RNNLM over n-gram increases with more data!

• Breakthrough result in 2011: 11% WER reduction over large system from IBM
• Ensemble of big RNNLM models trained on a lot of data

• RNNs became much more accessible through open-source
implementations in general ML toolkits:
• Theano
• Torch
• PyBrain
• TensorFlow
• …

Recurrent Nets Today
• Widely applied:
• ASR (both acoustic and language models)
• MT (language & translation & alignment models, joint models)
• Many NLP applications
• Video modeling, handwriting recognition, user intent prediction, …
• Downside: for many problems RNNs are too powerful, models are
becoming unnecessarily complex
• Often, complicated RNN architectures are preferred because of wrong
reasons (easier to get a paper published and attract attention)

Longer short term memory in simple RNNs
• How to add longer memory to RNNs without unnecessary complexity
• Paper: Learning Longer Memory in Recurrent Neural Networks
(Mikolov, Joulin, Chopra, Mathieu, Ranzato, ICLR Workshop 2015)

Recurrent Network – Elman Architecture
• Also known as Simple Recurrent Network (SRN)
• Input layer 𝑥 𝑡, hidden layer ℎ 𝑡, output 𝑦𝑡
• Weight matrices 𝐴, 𝑅, 𝑈

Recurrent Network – Elman Architecture
• Input layer 𝑥 𝑡, hidden layer ℎ 𝑡, output 𝑦𝑡
• Weight matrices 𝐴, 𝑅, 𝑈
ℎ 𝑡 = σ(𝐴𝑥 𝑡 + 𝑅ℎ 𝑡−1)
σ(𝑥) =
1
1 + 𝑒 𝑥
𝑦𝑡 = 𝑓(𝑈ℎ 𝑡)
𝑓() is softmax function

Simple Recurrent Net Problems
• Backpropagation through time algorithm + stochastic gradient
descent is commonly used for training (Rumelhart et al, 1985)
• Gradients can either vanish or explode (Hochreiter 1991;
Bengio 1994)

Simple Recurrent Net: Exploding Gradients
• The gradients explode rarely, but this can have disastrous effects
• Simple “hack” is to clip gradients to stay within some range
• This prevents exponential growth (which would later lead to giant
step in weight update)
• One can also normalize the gradients, or discard the weight updates
that are too big

Simple Recurrent Net: Vanishing Gradients
• Most of the time, the gradients quickly vanish (after 5-10 steps of
backpropagation through time)
• This may not be a problem of SGD, but of the architecture of the SRN

Simple Recurrent Net: Vanishing Gradients
• What recurrent architecture would be easier to train to capture
longer term patterns?
• Instead of fully connected recurrent matrix, we can use architecture
where each neuron is connected only to the input and to itself
• Old idea (Jordan 1987; Mozer 1989)

Combination of both ideas: Elman + Mozer
• Part of the hidden layer is fully connected,
part is diagonal (self-connections)
• Can be seen as RNN with two
hidden layers
• Or as RNN with partially diagonal
recurrent matrix (+ linear hidden units)

Combination of both ideas: Elman + Mozer
• The 𝛼 value can be learned, or kept
fixed close to 1 (we used 0.95)
• The 𝑃 matrix is optional
(usually helps a bit)

Structurally Constrained Recurrent Net
• Because we constrain the architecture of SRN, we further denote the
model as Structurally Constrained Recurrent Net (SCRN)
• Alternative name is “slow recurrent nets”, as the state of the diagonal
layer changes slowly
Q: Wouldn’t it be enough to initialize the recurrent matrix to be diagonal?
A: No. This would degrade back to normal RNN and not learn longer memory.

Results
• Language modeling experiments: Penn Treebank, Text8
• Longer memory in language models is commonly called cache / topic
• Comparison to Long Short Term Memory RNNs (currently popular but
quite complicated architecture that can learn longer term patterns)
• Datasets & code: http://github.com/facebook/SCRNNs
(link is in the paper)

Results: Penn Treebank language modeling
• Gain from SCRN / LSTM over simpler recurrent net is similar to gain from cache
• LSTM has 3 gates for each hidden unit, and thus 4x more parameters need to be
accessed during training for the given hidden layer size (=> slower to train)
• SCRN with 100 fully connected and 40 self-connected neurons is only slightly
more expensive to train than SRN
MODEL # hidden units Perplexity
N-gram - 141
N-gram + cache - 125
SRN 100 129
LSTM 100 (x4 parameters) 115
SCRN 100 + 40 115

Results: Text8
• Text8: Wikipedia text (~17M words), much stronger effect from cache
• Big gain for both SCRN & LSTM over SRN
• For small models, SCRN seems to be superior (simpler architecture, better
accuracy, faster training – less parameters)
N-gram - 309
SRN 100 245
SCRN 100 + 80 184

Results: Text8
• With 500 hidden units, LSTM is slightly better in perplexity (3%) than SCRN, but it
also has many more parameters
N-gram - 309
SRN 500 184
SCRN 500 + 80 161

Discussion of Results
• SCRN accumulates longer history in the “slow” hidden layer: the same
as exponentially decaying cache model
• Empirically, LSTM performance correlates strongly with cache
(weighted bag-of-words)
• For very large (~infinite) training sets, SCRN seems to be the
preferable architecture: it is computationally very cheap

Conclusion
• Simple tricks can overcome the vanishing and exploding gradient
problems
• State of the recurrent layer can represent longer short term memory,
but not the long term one (across millions of time steps)
• To represent true long term memory, we may need to develop models
with ability to grow in size (modify their own structure)

Beyond Deep Learning
• Going beyond: what RNNs and deep networks cannot model
efficiently?
• Surprisingly simple patterns! For example, memorization of
variable-length sequence of symbols

Beyond Deep Learning: Algorithmic Patterns
• Many complex patterns have short, finite description length in natural
language (or in any Turing-complete computational system)
• We call such patterns Algorithmic patterns
• Examples of algorithmic patterns: 𝑎 𝑛 𝑏 𝑛, sequence memorization,
addition of numbers learned from examples
• These patterns often cannot be learned with standard deep learning
techniques

Beyond Deep Learning: Algorithmic Patterns
• Among the myriad of complex tasks that are currently not solvable,
which ones should we focus on?
• We need to set ambitious end goal, and define a roadmap how to
achieve it step-by-step

A Roadmap towards
Machine Intelligence
Tomas Mikolov, Armand Joulin and Marco Baroni

Ultimate Goal for Communication-based AI
Can do almost anything:
• Machine that helps students to understand homeworks
• Help researchers to find relevant information
• Write programs
• Help scientists in tasks that are currently too demanding (would
require hundreds of years of work to solve)

The Roadmap
• We describe a minimal set of components we think the intelligent
machine will consist of
• Then, an approach to construct the machine
• And the requirements for the machine to be scalable

Components of Intelligent machines
• Ability to communicate
• Motivation component
• Learning skills (further requires long-term memory), ie. ability to
modify itself to adapt to new problems

Components of Framework
To build and develop intelligent machines, we need:
• An environment that can teach the machine basic communication skills and
learning strategies
• Communication channels
• Rewards
• Incremental structure

The need for new tasks: simulated
environment
• There is no existing dataset known to us that would allow to teach the
machine communication skills
• Careful design of the tasks, including how quickly the complexity is
growing, seems essential for success:
• If we add complexity too quickly, even correctly implemented intelligent
machine can fail to learn
• By adding complexity too slowly, we may miss the final goals

High-level description of the environment
Simulated environment:
• Learner
• Teacher
• Rewards
Scaling up:
• More complex tasks, less examples, less supervision
• Communication with real humans
• Real input signals (internet)

Simulated environment - agents
• Environment: simple script-based reactive agent that produces signals
for the learner, represents the world
• Learner: the intelligent machine which receives input signal, reward
signal and produces output signal to maximize average incoming
reward
• Teacher: specifies tasks for Learner, first based on scripts, later to be
replaced by human users

Simulated environment - communication
• Both Teacher and Environment write to Learner’s input channel
• Learner’s output channel influences its behavior in the Environment,
and can be used for communication with the Teacher
• Rewards are also part of the IO channels

Visualization for better understanding
• Example of input / output streams and visualization:

How to scale up: fast learners
• It is essential to develop fast learner: we can easily build a machine
today that will “solve” simple tasks in the simulated world using a
myriad of trials, but this will not scale to complex problems
• In general, showing the Learner new type of behavior and guiding it
through few tasks should be enough for it to generalize to similar
tasks later
• There should be less and less need for direct supervision through
rewards

How to scale up: adding humans
• Learner capable of fast learning can start communicating with human
experts (us) who will teach it novel behavior
• Later, a pre-trained Learner with basic communication skills can be
used by human non-experts

How to scale up: adding real world
• Learner can gain access to internet through its IO channels
• This can be done by teaching the Learner how to form a query in its
output stream

The need for new techniques
Certain trivial patterns are nowadays hard to learn:
• 𝑎 𝑛 𝑏 𝑛 context free language is out-of-scope of standard RNNs
• Sequence memorization breaks LSTM RNNs
• We show this in a recent paper Inferring Algorithmic Patterns with
Stack-Augmented Recurrent Nets

Scalability
To hope the machine can scale to more complex problems, we need:
• Long-term memory
• (Turing-) Complete and efficient computational model
• Incremental, compositional learning
• Fast learning from small number of examples
• Decreasing amount of supervision through rewards
• Further discussed in: A Roadmap towards Machine Intelligence
http://arxiv.org/abs/1511.08130

Some steps forward: Stack RNNs (Joulin &
Mikolov, 2015)
• Simple RNN extended with a long term memory module that the
neural net learns to control
• The idea itself is very old (from 80’s – 90’s)
• Our version is very simple and learns patterns with complexity far
exceeding what was shown before (though still very toyish): much
less supervision, scales to more complex tasks

• Learns algorithms from examples
• Add structured memory to RNN:
• Trainable [read/write]
• Unbounded
• Actions: PUSH / POP / NO-OP
• Examples of memory structures:
stacks, lists, queues, tapes, grids,
…
Stack RNN

Algorithmic Patterns
• Examples of simple algorithmic patterns generated by short programs
(grammars)
• The goal is to learn these patterns unsupervisedly just by observing the
example sequences

Algorithmic Patterns - Counting
• Performance on simple counting tasks
• RNN with sigmoidal activation function cannot count
• Stack-RNN and LSTM can count

Algorithmic Patterns - Sequences
• Sequence memorization and binary addition are out-of-scope of
LSTM
• Expandable memory of stacks allows to learn the solution

Binary Addition
• No supervision in training, just prediction
• Learns to: store digits, when to produce output, carry

Stack RNNs: summary
The good:
• Turing-complete model of computation (with >=2 stacks)
• Learns some algorithmic patterns
• Has long term memory
• Simple model that works for some problems that break RNNs and LSTMs
• Reproducible: https://github.com/facebook/Stack-RNN
The bad:
• The long term memory is used only to store partial computation (ie. learned skills are not
stored there yet)
• Does not seem to be a good model for incremental learning
• Stacks do not seem to be a very general choice for the topology of the memory

Conclusion
To achieve true artificial intelligence, we need:
• AI-complete goal
• New set of tasks
• Develop new techniques
• Motivate more people to address these problems

Recurrent networks and beyond by Tomas Mikolov

More Related Content

What's hot (20)

Viewers also liked (16)

Similar to Recurrent networks and beyond by Tomas Mikolov (20)

More from Bhaskar Mitra (20)

Recently uploaded (20)

Recurrent networks and beyond by Tomas Mikolov