Hierarchical Temporal Memory (HTM)

Jeff HawkinsMay 10, 2006IBM

A new computational paradigmbased on cortical theory

Pipe DreamDriven By

Greed

MotherOf All

Markets

Today’s PDAMarket Indicator

Not in ourLifetime

AnyMoment

Now

Today’s CognitiveComputing Indicator

Not in our lifetime- Decades of effort

- AI- neural networks- fuzzy logic- 5th generation project- decade of the brain

- Not much success- vision, language, robotics

- Brain is very complex

Not in our lifetime- Decades of effort

- AI- neural networks- fuzzy logic- 5th generation project- decade of the brain

- Not much success- vision, language, robotics

- Brain is very complex

Any moment now- Neocortex:

- Fast- Flexible- Robust

- 100 years of data- Anatomical, physiological- Mathematics

- Common cortical algorithm- Cortical Theory (HTM)

World HTM/CortexSenses

PeopleCarsBuildingsWordsSongsIdeas

patterns

World HTM/CortexSenses

PeopleCarsBuildingsWordsSongsIdeas

patterns

“Causes” “Beliefs”

cause1 0.22cause2 0.07cause3 0.00cause4 0.63cause5 0.00cause6 0.08

Causes Representations of Causes

HTM

What does an HTM do?

1 Discover causes in the world2 Infer causes of novel input3 Predict future4 Direct motor behavior

Sensory data

Belief

HTMs use a hierarchy of memory nodes

Sensory data

Beliefs

HTMs use a hierarchy of memory nodes

Each node: Discovers causes (of its input)Passes beliefs upPasses predictions down

Sensory data

Beliefs

HTMs use a hierarchy of memory nodes

Each node: Discovers causes (of its input)Passes beliefs upPasses predictions down

Each node: Stores common sequencesChanging sensory data forms stable beliefs at topStable beliefs at top form changing sensory predictions

1) Why does hierarchy make a difference?

2) How does each node discover and infer causes?

Why does hierarchy make a difference?

1) Shared representations lead to generalization and efficiency

Why does hierarchy make a difference?

1) Shared representations lead to generalization and efficiency

2) HTM hierarchy matches spatial and temporal hierarchy of causes in world

Why does hierarchy make a difference?

1) Shared representations lead to generalization and efficiency

2) HTM hierarchy matches spatial and temporal hierarchy of causes in world

3) Belief propagation techniques ensure all nodes quickly reach mutually compatible beliefs

80% woof20% meow

70% pig image30% cat image

90% cat

CPT

Belief Propagation

Why does hierarchy make a difference?

1) Shared representations lead to generalization and efficiency

2) HTM hierarchy matches spatial and temporal hierarchy of causes in world

3) Belief propagation techniques ensure all nodes quickly reach mutually compatible beliefs

4) Affords mechanism for attention

How does each node discover causes?

How does each node discover causes?

1) Learn common spatial patterns2) Learn common sequences of spatial patterns

How does each node discover causes?

1) Learn common spatial patterns

(things that happen at the same time are likely to have a common cause)

How does each node discover causes?

1) Learn common spatial patterns

Common patterns:remember

Uncommon patterns:ignore

How does each node discover causes?

1) Learn common spatial patterns2) Learn common sequences of spatial patterns

How does each node discover causes?

Common sequence:assign to cause

time

Common sequence:assign to cause

Uncommon sequence:ignore

1) Learn common spatial patterns2) Learn common sequences of spatial patterns

How does each node discover causes?

1) Learn common spatial patterns2) Learn common sequences3) Use context from above in hierarchy

Do HTMs really work?

4 pixels

Level 1

Level 2

Level 3

Simple HTM vision system (32x32 pixel)

Training images

Trainingimages

Correct Incorrect

Correctly recognized images

Numenta Plan1) Develop a detailed computational theory of

neocortical function (HTM) On Intelligence (Times Books, 2004) HTM white paper, www.numenta.com Biological mapping paper, August 2006

Numenta Plan1) Develop a detailed computational theory of

neocortical function (HTM)

2) Develop a software platform for HTM applications

Net list

Supervisor

Dev Tools

Configurator

Trainer

Debugger

SupervisorAPI

Node Processor 2

Node Processor N

:Gig

abit

sw

itch

Fileserver

NodeProcessor

Numenta Platform

Run time environment

Numenta Plan1) Develop a detailed computational theory of

neocortical function (HTM)

2) Develop a software platform for HTM applications

Multiple processor/server architecture Optimized C++ routines Developer toolset with flexible scripting using Python Supports Linux + MacOS. Windows to come. Build a community of developers

Early access partners, 2nd meeting end of May 2006 Beta release early 2007

Numenta Plan1) Develop a detailed computational theory of

neocortical function (HTM)

2) Develop a software platform for HTM applications

3) Test HTM with a machine vision system

Numenta Machine Vision System Robust Object Recognition From Natural Images Recognition Task Defined Data collection in process

Highly realistic 3D models and textures used to generate sequences 90,000 images and 102 sequences collected to date Each image has accurate alpha channel for programmatic 2D modifications

HTM Applications

1) What humans find easy and computers hard- vision, language, robotics- many apps from security to self-driving cars- extend with new senses, IR, sonar, radar…

2) Discovering causes in unusual worlds- geology, markets, weather, physics, genetics

HTM Capabilities

1) Discover causes2) Inference3) Prediction4) Behavior

Beyond biology

1) Faster2) Larger3) Exotic senses

www.OnIntelligence.org

www.Numenta.com(white paper posted this week)

Not in ourLifetime

AnyMoment

Now

Today’s CognitiveComputing Indicator

Thank _ _ _

HTM

HTM models world, including hardwired motor behaviors

world

motor

HTM

Representations of motor behavior are auto-associatively paired with motor generators

world

motor

Hierarchical Temporal Memory

1) Powerful, flexible, robust2) Can be applied to many problems

- vision- language- robotics- manufacturing- business modeling- market modeling- network modeling- resource exploration- weather prediction- math, physics

Discovering and inferring causes has proven to bevery difficult, e.g.

- visual pattern recognition- language understanding- machine learning

Sensory data

?

Beliefs(of causes)

“What is conspicuously lacking is a broad framework of ideas within which to interpret these different approaches.”

Francis Crick, 1979

Belief Propagation

Belief Propagation

“maybe diagonal line, maybe vertical line”

Belief Propagation

“maybe diagonal line, maybe vertical line”

“maybe diagonal line, maybe vertical line”