Modern Artificial Intelligence
S. M. Ali Eslami
December 2015
Chappie (2015)
Ex Machina (2015)
Elon Musk
Stephen Hawking
Outline
Output
Input
Algorithm
Programmable
Computer
“Horse”
Search
Output
Input
Algorithm
Programmable
Computer
Human
?
Image Classification
Output
Input
Algorithm
Horse
?
Computer
Output
Input
Algorithm
Horse
Image Classification
Tasks thought to require intelligence
Can a general system achieve all these tasks?
Immediate Usefulness
General Applicability
Immediate Usefulness
General Applicability
Immediate Usefulness
General Applicability
Immediate Usefulness
General Applicability
Immediate Usefulness
General Applicability
Immediate Usefulness
General Applicability
?
Immediate Usefulness
General Applicability
?
Deep Learning
?
Computer
Output
Input
Horse
Algorithm
Computer
Output
Input
Horse
Preprocessing
Feature Extraction
Feature Selection
Discrimination
Calibration
Algorithm
Computer
Horse
Cow
Output
Input
Preprocessing
Feature Extraction
Feature Selection
Learned
Discrimination
Calibration
Algorithm
Computer
Horse
Cow
Output
Input
Stage 1
Stage 2
Stage 3
Stage 4
Stage 5
Algorithm
Convolutional neural networks for image classification
Torch (2015)
Deep Learning
Krizhevsky et al. (2012)
Clarifai (2014)
Where do the labels come from?
Reinforcement Learning
Architecture
Agent
Environment
Observations
Actions
ATARI agents
100+ classic 8-bit Atari games
Human-level control through deep reinforcement learning
Mnih et al. (Nature, 2015)
Space Invaders agent
Breakout agent
General Atari agent
Human-level control through deep reinforcement learning
Mnih et al. (Nature, 2015)
Human-level control through deep reinforcement learning
Mnih et al. (Nature, 2015)
Deep Reinforcement Learning for Continuous Control
How many experiences do we need?
1 epoch = 50,000 interactions = 30 minutes of experience
Total experience: 10m interactions = 5 days
Model-based Methods
Three learning paradigms
x
z
x
Supervised
Learning
Reinforcement
Learning
y
z
a
horse
left
Three learning paradigms
Model
x
z
x
x
z
x
Supervised
Learning
Reinforcement
Learning
Generative
Modelling
y
z
a
y
horse
left
Three learning paradigms
Model
x
z
x
x
z
x
Supervised
Learning
Reinforcement
Learning
Generative
Modelling
y
z
a
y
(2.3, -1, 0.5, 3)
not blinking
horse
left
Learning to draw shapes
The Shape Boltzmann Machine
Sampling from an SBM
Learning to draw shapes
Learning to draw shapes
The Shape Boltzmann Machine: a Strong Model of Object Shape
S. M. Ali Eslami, Nicolas Heess, Christopher K. I. Williams, John Winn
International Journal of Computer Vision, Springer (IJCV, 2013)
Factored Shapes and Appearances
A Generative Model for Parts-based Object Segmentation
S. M. Ali Eslami, Christopher K. I. Williams
Neural Information Processing Systems (NIPS, 2012)
Learning to segment objects
A Generative Model for Parts-based Object Segmentation
S. M. Ali Eslami, Christopher K. I. Williams
Neural Information Processing Systems (NIPS, 2012)
Learning to segment objects
A Generative Model for Parts-based Object Segmentation
S. M. Ali Eslami, Christopher K. I. Williams
Neural Information Processing Systems (NIPS, 2012)
Recurrent Neural Networks for Image Generation
Gregor et al. (2015)
c
x
z
p(x|c)
Decoding
Generation
Encoding
Inference
Recurrent Neural Networks for Image Generation
Gregor et al. (2015)
c
x
z
p(x|c)
Decoding
Generation
Encoding
Inference
Write
Read
Read
ct
x
ct+1
x
Write
p(x|cT)
z
z
Recurrent Neural Networks for Image Generation
Gregor et al. (2015)
Recurrent Neural Networks for Image Generation
Gregor et al. (2015)
Generative Modelling
Model p(x|z) can be:
What models should we use?
Choice of model p(x|z)
almost always constrained
by our ability to compute p(z|x)
Model
x
z
x
y
How should we do inference?
Reinforced Variational Inference
Modern Variational Inference
Model
x
z
x
y
Consensus Message Passing for Layered Graphical Models
Varun Jampani, S. M. Ali Eslami, Daniel Tarlow Pushmeet Kohli, John Winn
Artificial Intelligence and Statistics (AISTATS, 2015)
p(z|x)
Modern Variational Inference
Model
x
z
x
y
Consensus Message Passing for Layered Graphical Models
Varun Jampani, S. M. Ali Eslami, Daniel Tarlow Pushmeet Kohli, John Winn
Artificial Intelligence and Statistics (AISTATS, 2015)
p(z|x)
q0(z)
Modern Variational Inference
Model
x
z
x
y
Consensus Message Passing for Layered Graphical Models
Varun Jampani, S. M. Ali Eslami, Daniel Tarlow Pushmeet Kohli, John Winn
Artificial Intelligence and Statistics (AISTATS, 2015)
p(z|x)
q0(z)
q1(z)
Modern Variational Inference
Model
x
z
x
y
Consensus Message Passing for Layered Graphical Models
Varun Jampani, S. M. Ali Eslami, Daniel Tarlow Pushmeet Kohli, John Winn
Artificial Intelligence and Statistics (AISTATS, 2015)
p(z|x)
q0(z)
q1(z)
q(z|x)
x
Modern Variational Inference
as codes representing the image x
Model
x
z
x
y
Consensus Message Passing for Layered Graphical Models
Varun Jampani, S. M. Ali Eslami, Daniel Tarlow Pushmeet Kohli, John Winn
Artificial Intelligence and Statistics (AISTATS, 2015)
Minimising the KL
by many (e.g. VAE, DLGM, NVIL, etc.)
Model
x
z
x
y
Variational Inference as Reinforcement Learning
Reinforced Variational Inference
Theophane Weber, Nicolas Heess, S. M. Ali Eslami, John Schulman, David Wingate, David Silver. Neural Information Processing Systems, Workshop on Advances in Approximate Bayesian Inference (NIPS, 2015)
maximise
Summary
Prediction as a subset of inference
Inference as a reinforcement learning problem
Reinforcement learning as a deep learning problem
http://arkitus.com
ali@arkitus.com