Large-Scale Deep Learning for

Intelligent Computer Systems

Jeff Dean
Google Brain team in collaboration with many other teams

Growing Use of Deep Learning at Google

# of directories containing model description files

Across many
products/areas:
Android
Apps
GMail
Image Understanding
Maps
NLP
Photos
Robotics
Speech
Translation
many research uses..
YouTube
many others ...

Outline
Two generations of deep learning software systems:
1st generation: DistBelief [Dean et al., NIPS 2012]
2nd generation: TensorFlow (unpublished)
An overview of how we use these in research and products
Plus, ...a new approach for training (people, not models)

Google Brain project started in 2011, with a focus on

pushing state-of-the-art in neural networks. Initial
emphasis:
use large datasets, and
large amounts of computation
to push boundaries of what is possible in perception and
language understanding

Plenty of raw data

Text: trillions of words of English + other languages

Visual data: billions of images and videos
Audio: tens of thousands of hours of speech per day
User activity: queries, marking messages spam, etc.
Knowledge graph: billions of labelled relation triples
...

How can we build systems that truly understand this data?

Text Understanding
This movie should have NEVER been made. From the poorly
done animation, to the beyond bad acting. I am not sure at what
point the people behind this movie said "Ok, looks good! Lets
do it!" I was in awe of how truly horrid this movie was.

Turnaround Time and Effect on Research

Minutes, Hours:

Interactive research! Instant gratification!

1-4 days

Tolerable
Interactivity replaced by running many experiments in parallel

1-4 weeks:

High value experiments only

Progress stalls

>1 month

Dont even try

Important Property of Neural Networks

Results get better with

more data +
bigger models +
more computation
(Better algorithms, new insights and improved
techniques always help, too!)

How Can We Train Large, Powerful Models Quickly?

Exploit many kinds of parallelism
Model parallelism
Data parallelism

Model Parallelism

Model Parallelism

Model Parallelism

Data Parallelism
Parameter Servers

Model
Replicas

...

Data

...

Data Parallelism
Parameter Servers

p
Model
Replicas

...

Data

...

Data Parallelism
Parameter Servers

Model
Replicas

...

Data

...

Data Parallelism
Parameter Servers

p = p + p

Model
Replicas

...

Data

...

Data Parallelism
Parameter Servers

p = p + p

p
Model
Replicas

...

Data

...

Data Parallelism
Parameter Servers

Model
Replicas

...

Data

...

Data Parallelism
Parameter Servers

p = p + p

Model
Replicas

...

Data

...

Data Parallelism
Parameter Servers

p = p + p

Model
Replicas

...

Data

...

Data Parallelism Choices

Can do this synchronously:

N replicas eqivalent to an N times larger batch size

Pro: No noise
Con: Less fault tolerant (requires recovery if any single machine fails)

Can do this asynchronously:

Con: Noise in gradients

Pro: Relatively fault tolerant (failure in model replica doesnt block other
replicas)

(Or hybrid: M asynchronous groups of N synchronous replicas)

Data Parallelism Considerations

Want model computation time to be large relative to time to send/receive
parameters over network
Models with fewer parameters, that reuse each parameter multiple times in the
computation

Mini-batches of size B reuse parameters B times

Certain model structures reuse parameter many times within each example:

Convolutional models tend to reuse hundreds or thousands of times per

example (for different spatial positions)
Recurrent models (LSTMs, RNNs) tend to reuse tens to hundreds of times
(for unrolling through T time steps during training)

What are some ways that

deep learning is having
a significant impact at Google?

Sequence to Sequence Models

Oriol Vinyals, Ilya Sutskever & Quoc Le started looking at how to map one
sequence to another sequence:
0.1
0.5
1.0
0.0
2.4

The lazy brown fox jumped over the ...

Sequence to Sequence Model

Deep LSTM

High dimensional
representation of a
sequence
0.1
0.5
1.0
0.0
2.4

The lazy brown fox jumped over the ...

Connect two, you get a machine translation

system
Le renard brun rapide saute par dessus ...

The quick brown fox jumped over the ...

It works well
WMT14

BLEU

State-of-the-art

37.0

Neural Translation Model

37.3

Sequence to Sequence Learning with Neural Networks

Ilya Sutskever, Oriol Vinyals, Quoc V. Le (NIPS 2014)
Addressing Rare Word Problems in Neural Translation Models (arxiv.org/abs/1410.8206)
Thang Luong, Ilya Sutskever, Oriol Vinyals, Quoc V. Le, Wojciech Zaremba

 or a chatbot.
I'm fine, thank you!

Hi Shiva, how are you?

It produces sensible conversations.

User> hello , i have a problem with my machine
BrainStop> hi
BrainStop> this is shiva .
BrainStop> how are you doing today ?
User> hi shiva , how are you ?
BrainStop> i ' m fine , thank you !
BrainStop> how may i assist you today ?
User> i want to access using vpn
BrainStop> currently the computer is connected to corp network ?
User> no
BrainStop> check this solution :
BrainStop> https : / / INTERNAL_HOSTNAME_REDACTED / articles / connecting - to disconnecting - from - ssl - vpn # heading - 4
User> thanks , bye !
BrainStop> thank you .

 or a parser.
n:(S.17 n:(S.17 n:(NP.11 p:NNP.53 n:) ...

Allen is locked in, regardless of his situ...

It works well
Completely learned parser with no parsing-specific code
State of the art results on WSJ 23 parsing task
Grammar as a Foreign Language, Oriol Vinyals, Lukasz Kaiser, Terry Koo, Slav
Petrov, Ilya Sutskever, and Geoffrey Hinton (to appear in NIPS 2015)
http://arxiv.org/abs/1412.7449

 or something that can learn graph algorithms

output:
Convex Hull
(or Delauney
Triangulation)
(or Travelling
Salesman tour)

input:
collection
of points

Pointer Networks, Oriol Vinyals, Meire Fortunato, &

Navdeep Jaitly (to appear in NIPS 2015)

Object Recognition Improvement Over Time

Predicted Human Performance

cat

ImageNet Challenge Winners

Image Models

cat

Module with 6
separate
=
convolutional
layers
24 layers deep

Going Deeper with Convolutions

Szegedy et al. CVPR 2015

Good Fine-Grained Classification

Good Generalization

Both recognized as meal

Sensible Errors

Works in practice for real users

Works in practice for real users

Connect sequence and image models, you get a

captioning system
A close up of a child holding a stuffed animal

It works well (BLEU scores)

Dataset

Previous SOTA

Show & Tell

Human

MS COCO

N/A

67

69

FLICKR

49

63

68

PASCAL (xfer learning)

25

59

68

SBU (weak label)

11

27

N/A

Show and Tell: A Neural Image Caption Generator,

Oriol Vinyals, Alexander Toshev, Samy Bengio, and Dumitru Erhan (CVPR
2015)

TensorFlow:
Second Generation Deep Learning System

Motivations
DistBelief (1st system) was great for scalability
Not as flexible as we wanted for research purposes
Better understanding of problem space allowed us to
make some dramatic simplifications

TensorFlow: Expressing High-Level ML Computations

Core in C++
Very low overhead
Different front ends for specifying/driving the computation
Python and C++ today, easy to add more

...

Python front end

C++ front end

Core TensorFlow Execution System

CPU

GPU

Android

iOS

...

TensorFlow Example (Batch Logistic Regression)

graph = tf.Graph()
with graph.AsDefault():
examples = tf.constant(train_dataset)
labels = tf.constant(train_labels)
W = tf.Variable(tf.truncated_normal([image_size * image_size, num_labels]))
b = tf.Variable(tf.zeros([num_labels]))

# Create new computation graph

# Training data/labels

# Variables

logits = tf.mat_mul(examples, W) + b
# Training computation
loss = tf.reduce_mean(tf.nn.softmax_cross_entropy_with_logits(logits, labels))
optimizer = tf.train.GradientDescentOptimizer(0.5).minimize(loss)
prediction = tf.nn.softmax(logits)

# Optimizer to use
# Predictions for training data

TensorFlow Example (Batch Logistic Regression)

graph = tf.Graph()
with graph.AsDefault():
examples = tf.constant(train_dataset)
labels = tf.constant(train_labels)
W = tf.Variable(tf.truncated_normal([image_size * image_size, num_labels]))
b = tf.Variable(tf.zeros([num_labels]))

# Create new computation graph

# Training data/labels

# Variables

logits = tf.mat_mul(examples, W) + b
# Training computation
loss = tf.reduce_mean(tf.nn.softmax_cross_entropy_with_logits(logits, labels))
optimizer = tf.train.GradientDescentOptimizer(0.5).minimize(loss)
prediction = tf.nn.softmax(logits)
with tf.Session(graph=graph) as session:
tf.InitializeAllVariables().Run()
for step in xrange(num_steps):
_, l, predictions = session.Run([optimizer, loss, prediction])
if (step % 100 == 0):
print 'Loss at step', step, ':', l
print 'Training accuracy: %.1f%%' % accuracy(predictions, labels)

# Optimizer to use
# Predictions for training data

# Run & return 3 values

Computation is a dataflow graph

Graph of Nodes, also called Operations or ops.

biases

Add

weights
MatMul
examples

labels

Relu
Xent

Computation is a dataflow graph

Edges are N-dimensional arrays: Tensors

biases

Add

weights
MatMul
examples

labels

with

s
r
o
s
ten

Relu
Xent

Computation is a dataflow graph

'Biases' is a variable

e
t
a
t
ith s

Some ops compute gradients

= updates biases

biases

...

learning rate

Add

...

Mul

Computation is a dataflow graph

Device A

biases

...

d
e
t
u
b
i
r
t
is

Add

learning rate

Devices: Processes, Machines, GPUs, etc

...

Mul

Device B

TensorFlow: Expressing High-Level ML Computations

Automatically runs models on range of platforms:

from phones ...

to single machines (CPU and/or GPUs)

to distributed systems of many 100s of GPU cards

What is in a name?
Tensor: N-dimensional array

1-dimension: Vector
2-dimension: Matrix
Represent many dimensional data flowing through the graph
e.g. Image represented as 3-d tensor rows, cols, color

Flow: Computation based on data flow graphs

Lots of operations (nodes in the graph) applied to data flowing through

Tensors flow through the graph TensorFlow

Edges represent the tensors (data)

Nodes represent the processing

Flexible
General computational infrastructure
Deep Learning support is a set of libraries on top of the core
Also useful for other machine learning algorithms
Possibly even for high performance computing (HPC) work
Abstracts away the underlying devices/computational hardware

Extensible
Core system defines a number of standard operations
and kernels (device-specific implementations of
operations)
Easy to define new operators and/or kernels

Deep Learning in TensorFlow

Typical neural net layer maps to one or more tensor operations

e.g. Hidden Layer: activations = Relu(weights * inputs + biases)

Library of operations specialized for Deep Learning

Dozens of high-level operations: 2D and 3D convolutions, Pooling, Softmax, ...

Standard losses e.g. CrossEntropy, L1, L2

Various optimizers e.g. Gradient Descent, AdaGrad, L-BFGS, ...

Auto Differentiation

Easy to experiment with (or combine!) a wide variety of different models:

LSTMs, convolutional models, attention models, reinforcement learning,
embedding models, Neural Turing Machine-like models, ...

No distinct Parameter Server subsystem

Parameters are now just stateful nodes in the graph
Data parallel training just a more complex graph
update

model
computation

update

model
computation

parameters

update

model
computation

Synchronous Variant
update

add

gradient

model
computation

gradient

model
computation

parameters

gradient

model
computation

Nurturing Great Researchers

Were always looking for people with the potential to become excellent
machine learning researchers

The resurgence of deep learning in the last few years has caused a surge of
interest of people who want to learn more and conduct research in this area

Google Brain Residency Program

New one year immersion program in deep learning research
Learn to conduct deep learning research w/experts in our team

Fixed one-year employment with salary, benefits, ...

Goal after one year is to have conducted several research projects

Interesting problems, TensorFlow, and access to computational resources

Google Brain Residency Program

Who should apply?

people with BSc or MSc, ideally in computer science, mathematics or statistics

completed coursework in calculus, linear algebra, and probability, or equiv.

programming experience

motivated, hard working, and have a strong interest in Deep Learning

Google Brain Residency Program

Program Application & Timeline

Google Brain Residency Program

For more information:

g.co/brainresidency
Contact us:

[email protected]

Questions?