What Deep Learning Means for Artificial Intelligence

What Deep Learning Means for
Artificial Intelligence
Jonathan Mugan, PhD
Encore Presentation to Austin Data Geeks
November 1, 2016

AI through the lens of System 1 and System 2
Psychologist Daniel Kahneman in Thinking Fast and Slow describes
humans as having two modes of thought: System 1 and System 2.
System 1: Fast and Parallel System 2: Slow and Serial
AI systems in these domains are useful
but limited. Called GOFAI (Good, Old-
Fashioned Artificial Intelligence).
1. Search and planning
2. Logic
3. Rule-based systems
AI systems in these domains have
been lacking.
1. Serial computers too slow
2. Lack of training data
3. Didn’t have the right algorithms
Subconscious: E.g., face recognition or
speech understanding.
Conscious: E.g., when listening to a
conversation or making PowerPoint slides.
We underestimated how hard it would
be to implement. E.g., we thought
computer vision would be easy.
We assumed it was the most difficult.
E.g., we thought chess was hard.

AI through the lens of System 1 and System 2
Psychologist Daniel Kahneman in Thinking Fast and Slow describes
humans as having two modes of thought: System 1 and System 2.
System 1: Fast and Parallel System 2: Slow and Serial
AI systems in these domains are useful
but limited. Called GOFAI (Good, Old-
Fashioned Artificial Intelligence).
1. Search and planning
2. Logic
3. Rule-based systems
Subconscious: E.g., face recognition or
speech understanding.
Conscious: E.g., when listening to a
conversation or making PowerPoint slides.
We underestimated how hard it would
be to implement. E.g., we thought
computer vision would be easy.
We assumed it was the most difficult.
E.g., we thought chess was hard.
This has changed
1. We now have GPUs and
distributed computing
2. We have Big Data
3. We have new algorithms [Bengio et
al., 2003; Hinton et al., 2006; Ranzato et
al., 2006]

Deep learning begins with a little function
It all starts with a humble linear function called a perceptron.
weight1 ✖ input1
weight2 ✖ input2
weight3 ✖ input3
sum
✚
Perceptron:
If sum > threshold: output 1
Else: output 0
Example: The inputs can be your data. Question: Should I buy this car?
0.2 ✖ gas mileage
0.3 ✖ horsepower
0.5 ✖ num cup holders
sum
✚
Perceptron:
If sum > threshold: buy
Else: walk

These little functions are chained together
Deep learning comes from chaining a bunch of these little
functions together. Chained together, they are called
neurons.
To create a neuron, we add a nonlinearity to the perceptron to get
extra representational power when we chain them together.
Our nonlinear perceptron is
sometimes called a sigmoid.
Plot of a sigmoid
where
The value b just offsets the sigmoid so the
center is at 0.

Single artificial neuron
Output, or input to
next neuron
weight1 ✖ input1
weight2 ✖ input2
weight3 ✖ input3

Three-layered neural network
A bunch of neurons chained together is called a neural network.
Layer 2: hidden layer. Called
this because it is neither input
nor output.
Layer 3: output. E.g., cat
or not a cat; buy the car or
walk.
Layer 1: input data. Can
be pixel values or the number
of cup holders.
This network has three layers.
(Some edges lighter
to avoid clutter.)
[16.2, 17.3, −52.3, 11.1]

Training with supervised learning
Supervised Learning: You show the network a bunch of things
with a labels saying what they are, and you want the network to
learn to classify future things without labels.
Example: here are some
pictures of cats. Tell me
which of these other pictures
are of cats.
To train the network, want to
find the weights that
correctly classify all of the
training examples. You hope
it will work on the testing
examples.
Done with an algorithm
called Backpropagation
[Rumelhart et al., 1986].
[16.2, 17.3, −52.3, 11.1]

Training with supervised learning
Supervised Learning: You show the network a bunch of things
with a labels saying what they are, and you want the network to
learn to classify future things without labels.
𝑤
𝑊
𝑦
𝑥 [16.2, 17.3, −52.3, 11.1]
Why Google’s Deep Learning toolbox is called TensorFlow.
y: output
x: input
h: number of hidden neurons
n: length of vector x
Learning is learning parameter
values

Deep learning is adding more layers
There is no exact
definition of
what constitutes
“deep learning.”
The number of
weights (parameters)
is generally large.
Some networks have
millions of parameters
that are learned.
(Some edges omitted
to avoid clutter.)
[16.2, 17.3, −52.3, 11.1]

Deep learning is adding more layers
Many “paths” for
updates in weights to
effect error.
Backpropagation is
like dynamic
programming.
“Chain Rule + Dynamic
Programming = Neural
Networks”
Edward Z. Yang
http://blog.ezyang.com/2011/05/neural-
networks
(Some edges omitted
to avoid clutter.)
[16.2, 17.3, −52.3, 11.1]

Talk Outline
• Introduction
• Deep learning and natural language processing
• Deep learning and computer vision
• Deep learning and robot actions
• What deep learning still can’t do
• Practical ways you can get started
• Conclusion
• About DeepGrammar (4 minutes, if time)

Deep learning enables sub-symbolic processing
Symbolic systems can be brittle.
I
bought
a
car
.
<i>
<bought>
<a>
<car>
<.>
You have to remember to
represent “purchased” and
“automobile.”
What about “truck”?
How do you encode the
meaning of the entire
sentence?

Recall our standard architecture
Layer 2: hidden layer. Called
this because it is neither input
nor output.
Layer 3: output. E.g., cat
or not a cat; buy the car or
walk.
Layer 1: input data. Can
be pixel values or the number
of cup holders.
Is this a cat?
[16.2, 17.3, −52.3, 11.1]

Neural nets with multiple outputs
Okay, but what kind of cat is it?
𝑃(𝑥)𝑃(𝑥)𝑃(𝑥) 𝑃(𝑥) 𝑃(𝑥)
Introduce a new node
called a softmax.
Probability a
house cat
Probability a
lion
Probability a
panther
Probability a
bobcat
Just normalize the
output over the sum of
the other outputs
(using the exponential).
Gives a probability.
[16.2, 17.3, −52.3, 11.1]

Learning word vectors
13.2, 5.4, −3.1 [−12.1, 13.1, 0.1] [7.2, 3.2,-1.9]
the man ran
From the sentence, “The man ran fast.”
𝑃(𝑥)𝑃(𝑥) 𝑃(𝑥) 𝑃(𝑥)
Probability
of “fast”
Probability
of “slow”
Probability
of “taco”
Probability
of “bobcat”
Learns a vector for each word based on the “meaning” in the sentence by
trying to predict the next word [Bengio et al., 2003].
These numbers updated
along with the weights
and become the vector
representations of the
words.

Comparing vector and symbolic representations
Vector representation
taco = [17.32, 82.9, −4.6, 7.2]
Symbolic representation
taco = 𝑡𝑎𝑐𝑜
• Vectors have a similarity score.
• A taco is not a burrito but similar.
• Symbols can be the same or not.
• A taco is just as different from a
burrito as a Toyota.
• Vectors have internal structure
[Mikolov et al., 2013].
• Italy – Rome = France – Paris
• King – Queen = Man – Woman
• Symbols have no structure.
• Symbols are arbitrarily assigned.
• Meaning relative to other symbols.
• Vectors are grounded in
experience.
• Meaning relative to predictions.
• Ability to learn representations
makes agents less brittle.

Yeah, that’s a word
But what about a sentence?
Algorithm for generating vectors for sentences
1. Make the sentence vector be the vector for the first
word.
2. For each subsequent word, combine its vector with the
sentence vector.
3. The resulting vector after the last word is the sentence
vector.
Can be implemented using a recurrent neural
network (RNN)

Encoding sentence meaning into a vector
h0
The
“The patient fell.”

h0
The
h1
patient

h0
The
h1
patient
h2
fell

Like a hidden Markov model, but doesn’t make the Markov
assumption and benefits from a vector representation.
h0
The
h1
patient
h2
fell
h3
.

Cool, a sentence vector
But what can you do with it?
You can unwind in the other direction to do
machine translation.
Called a seq2seq model, or Neural Machine
Translation, or encoder-decoder model.
You can feed it to a classifier.

Decoding sentence meaning
Machine translation, or structure learning more generally.
El
h3

El
h3 h4

El
h3
paciente
h4

El
h3
paciente
h4
cayó
h5
.
h6
[Cho et al., 2014]
It keeps generating until it generates a stop symbol.

Generating image captions
Convolutional
neural network
An
h0
angry
h1
sister
h2
.
h3
[Karpathy and Fei-Fei, 2015]
[Vinyals et al., 2015]Image

Image caption examples
[Karpathy and Fei-Fei, 2015] http://cs.stanford.edu/people/karpathy/deepimagesent/
See:

Attention [Bahdanau et al., 2014]
El
h3
paciente
h4
cayó
h5
.
h6
h0
The
h1
patient
h2
fell
h3
.

RNNs and Structure Learning
• In addition to machine translation and generating
captions for images, can be used to learn just about
any kind of structure you’d want, as long as you
have lots of training data.

Deep learning and question answering
RNNs answer questions.
What is the translation of this
phrase to French?
What is the next word?
Attention is useful for question
answering.
This can be generalized to which facts
the learner should pay attention to
when answering questions.

Deep learning and question answering
Bob went home.
Tim went to the junkyard.
Bob picked up the jar.
Bob went to town.
Where is the jar? A: town
• Memory Networks [Weston et al.,
2014]
• Updates memory vectors based on
a question and finds the best one to
give the output.
The office is north of the yard.
The bath is north of the office.
The yard is west of the kitchen.
How do you go from the office to
the kitchen? A: south, east
• Neural Reasoner [Peng et al.,
2015]
• Encodes the question and facts in
many layers, and the final layer is
put through a function that gives
the answer.

Vision is hard
• Even harder for 3D
objects.
• You move a bit, and
everything changes.
How a computer sees an image
[22, 81, 44, 88, 17, 0, ..., 45]
Example from MNIST
handwritten digit dataset
[LeCun and Cortes, 1998].
Vision is hard because images are big matrices of numbers.

Breakthrough: Unsupervised Model
• Big breakthrough in 2006 by Hinton et al.
• Use a network with symmetric weights called a restricted Boltzmann machine.
• Stochastic binary neuron.
• Probabilistically outputs 0
(turns off) or 1 (turns on)
based on the weight of the
inputs from on units.
0 1
• Limit connections to be from one
layer to the next.
• Fast because decisions are made
locally.
• Trained in an unsupervised way
to reproduce the data.
0 1 0 1 0 1
0 1 0 1 0 1

Stack up the layers to make a deep network
Input data
Hidden
The output of each layer
becomes the input to the
next layer [Hinton et al.,
2006].
See video starting at
second 45
https://www.coursera.org
/course/neuralnets
0 1 0 1 0 1
0 1 0 1 0 1
Input data
Hidden
0 1 0 1 0 1
0 1 0 1 0 1
Hidden layer becomes
input data of next layer.

Computer vision, scaling up
Unsupervised learning was scaled up
by Honglak Lee et al. [2009] to learn
high-level visual features.
[Lee et al., 2009]
Further scaled up by Quoc Le et al.
[2012].
• Used 1,000 machines (16,000
cores) running for 3 days to train 1
billion weights by watching
YouTube videos.
• The network learned to identify
cats.
• The network wasn’t told to look
for cats, it naturally learned that
cats were integral to online
viewing.
• Video on the topic at NYT
http://www.nytimes.com/2012/06/26/tec
hnology/in-a-big-network-of-computers-
evidence-of-machine-learning.html

Why is this significant?
To have a grounded understanding
of its environment, an agent must
be able to acquire representations
through experience [Pierce et al.,
1997; Mugan et al., 2012].
Without a grounded understanding,
the agent is limited to what was
programmed in.
We saw that unsupervised learning
could be used to learn the meanings of words, grounded in the
experience of reading.
Using these deep Boltzmann machines, machines can learn
to see the world through experience.

CNNs: Limit connections and duplicate parameters
Convolutional neural networks (CNNs)
build in a kind of feature invariance.
1. Convolution layers
• Bank of feature detectors
• Different feature detectors slide
over the image
2. Sub-sampling layers
• The next layer pools from a
region on the layer below With the layers and
topology, our networks are
starting to look a little like
the visual cortex. Although,
we still don’t fully
understand the visual
cortex.
Modern image processing
systems have many such layer
pairs.

More recent deep vision networks
ImageNet http://www.image-net.org/ is a huge collection of images
corresponding to the nouns of the WordNet hierarchy. There are hundreds to
thousands of images per noun.
2012 – Deep Learning begins to dominate image recognition
Krizhevsky et al. [2012] got 16% error on recognizing objects, when
before the best error was 26%. They used a convolutional neural
network (CNN).
2015 – Deep Learning surpasses human level performance
He et al. [2015] surpassed human level performance on recognizing
images of objects.* Computers seem to have an advantage when the
classes of objects are fine grained, such as multiple species of dogs.
*But deep learning can be easily fooled [Nguyen et al., 2014]. Enlightening video at
https://www.youtube.com/watch?v=M2IebCN9Ht4.

A stamping in of behavior
When we think of doing things, we think of conscious planning
with System 2.
Imagine trying to get to Seattle.
• Get to the airport. How? Take a taxi. How? Call a taxi. How?
Find my phone.
• Some behaviors arise more from a
a gradual stamping in [Thorndike,
1898].
• Became the study of Behaviorism
[Skinner, 1953] (see Skinner box
on the right).
• Formulated into artificial
intelligence as Reinforcement
Learning [Sutton and Barto, 1998]. By Andreas1 (Adapted from Image:Boite skinner.jpg) [GFDL
(http://www.gnu.org/copyleft/fdl.html) or CC-BY-SA-3.0
(http://creativecommons.org/licenses/by-sa/3.0/)], via
Wikimedia Commons
A “Skinner box”

Beginning with random exploration
In reinforcement learning, the agent begins by
randomly exploring until it reaches its goal.

Reaching the goal
• When it reaches the goal, credit is propagated back to its previous states.
• The agent learns the function Q(s,a), which gives the cumulative expected
discounted reward of being in state s and taking action a and acting according
to the policy thereafter.

Eventually, the agent learns the value of being in each state and
taking each action and can therefore always do the best thing in
each state.
Learning the behavior

Playing Atari with deep learning
Input, last four frames, where each frame
is downsampled to 84 by 84 pixels.
[Mnih et al., 2013]
represent the
state-action value
function as a
convolutional
neural network.
𝑃(𝑥)𝑃(𝑥) 𝑃(𝑥) 𝑃(𝑥)
Value of
moving left
Value of
moving right
Value of
shooting
Value of
reloading
In [Mnih et al., 2013],
this is actually three
hidden layers.
See some videos at
http://mashable.com/2015/02
/25/computer-wins-at-atari-
games/

What deep learning still can’t do
System 1: Fast and Parallel
• With computer vision, we seem to be on the
right track
• Reinforcement learning is useful in increasingly
large worlds
System 2: Slow and Serial
• Still lacking in common sense
• Language processing needs a grounded
understanding

Limitations of deep learning
The encoded meaning is grounded with respect to other words.
There is no linkage to the physical world.
"ICubLugan01 Reaching". Licensed under CC BY-SA 3.0 via Wikipedia - https://en.wikipedia.org/wiki/File:ICubLugan01_Reaching.png#/media/File:ICubLugan01_Reaching.png
The iCub http://www.icub.org/

Bob went home.
Tim went to the junkyard.
Bob picked up the jar.
Bob went to town.
Where is the jar? A: town
Deep learning has no
understanding of what it means for
the jar to be in town.
For example that it can’t also be at
the junkyard. Or that it may be in
Bob’s car, or still in his hands.
The encoded meaning is grounded with respect to other words.
There is no linkage to the physical world.

Imagine a dude standing
on a table. How would a
computer know that if
you move the table you
also move the dude?
Likewise, how could a
computer know that it
only rains outside?
Or, as Marvin Minsky asks, how could a computer learn
that you can pull a box with a string but not push it?

Imagine a dude standing
on a table. How would a
computer know that if
you move the table you
also move the dude?
Likewise, how could a
computer know that it
only rains outside?
Or, as Marvin Minsky asks, how could a computer learn
that you can pull a box with a string but not push it?
No one knows how to explain
all of these situations to a
computer. There’s just too
many variations.
A robot can learn through
experience, but it must be
able to efficiently generalize
that experience.

Best learning resources
Hinton’s Coursera Course. Get it right from the horse’s mouth.
He explains things well.
https://www.coursera.org/course/neuralnets
Online textbook in preparation for deep learning from Yoshua
Bengio and friends. Clear and understandable.
http://www.deeplearningbook.org/
TensorFlow tutorials.
https://www.tensorflow.org/versions/r0.11/tutorials/index.ht
ml
Deep Learning by Google
https://www.udacity.com/course/deep-learning--ud730

Conclusion
System 1: Fast and Parallel
Before deep learning, this was the strength of AI
With deep learning, this is now our strength
Time to get back to work on this

The importance of finding dumb mistakes

Thanks for listening
Jonathan Mugan
@jmugan
www.deepgrammar.com

What Deep Learning Means for Artificial Intelligence

More Related Content

What's hot (20)

Viewers also liked (20)

Similar to What Deep Learning Means for Artificial Intelligence (20)

More from Jonathan Mugan (6)

Recently uploaded (20)

What Deep Learning Means for Artificial Intelligence