Computing Like the Brain

Computing Like the Brain

Computing Like the Brain The Path To Machine Intelligence Jeff Hawkins GROK - Numenta [email protected] 1) Discover operating principles of...

2MB Sizes 0 Downloads 17 Views

Recommend Documents

Corrado Greg_Google Brain - The Brain Forum
What is the Google Brain Team? Started in 2011 as a 3-person project to explore large-scale training of artificial neura

His Brain, Her Brain
Apr 25, 2005 - These inequities are not just interesting idiosyncrasies that might explain why more men than women enjoy

BRAIN:
Batten, Frederick E. Epidemiology of poliomyelitis ... Ladd, G. T., and Woodworth, R S. Elements of ... Woodworth, R. S.

Language and the Brain
In addition ould complete certain phrases. For example, if she heard “Roses are red,” she would. Roses are red, viol

the human brain - AABGU
expense, the Glazers provided. 95 tons of medicine and special food to China in the earthquake area. DONOR IMPACT. DIANE

The Female Brain - rNawrocki
a mellowing agent; mother of all allopregnenolone (the brain's Valium, i.e., chill pill). ..... females who orgasm after

Brain Science - European Brain Council
Mar 3, 2015 - Lead the way to advance science, technology and health. The original Non Solus (not alone) mark of Isaac E

Computing in the national curriculum - Computing At School
but perhaps unfamiliar language used in the programme .... for the top and one for the bottom. ... Use two or more progr

Computing History D - The Centre for Computing History
The first 'Home Computer' was the Altair 8800. This machine was unveiled in January. 1975 on the cover of Popular Electr

Computing Like the Brain The Path To Machine Intelligence Jeff Hawkins GROK - Numenta [email protected]

1) Discover operating principles of neocortex

2) Build systems based on these principles

Artificial Intelligence - no neuroscience Alan Turing

“Computers are universal machines” “Human behavior as test for machine intelligence”

1935 1950

Major AI Initiatives • • • •

MIT AI Lab 5th Generation Computing Project DARPA Strategic Computing Initiative DARPA Grand Challenge

AI Projects • ACT-R • Asimo • CoJACK • Cyc • Deep Blue • Global Workspace Theory • Mycin • SHRDLU • Soar • Watson - Many more -

Pros:

- Good solutions

Cons:

- Task specific - Limited or no learning

Artificial Neural Networks – minimal neuroscience Warren McCulloch Walter Pitts

“Neurons as logic gates” 1943 Proposed first artificial neural network

ANN techniques • • • • •

Back propagation Boltzman machines Hopfield networks Kohonen networks Parallel Distributed Processing

• Machine learning • Deep Learning

Pros:

- Good classifiers - Learning systems

Cons:

- Limited - Not brain like

Whole Brain Simulator – maximal neuroscience The Human Brain Project

Blue Brain simulation

No theory No attempt at Machine Intelligence

1) Discover operating principles of neocortex 2) Build systems based on these principles

Anatomy, Physiology

Theory

Software

Silicon

Good progress is being made 1940s in computing = 2010s in machine intelligence

The neocortex is a memory system.

retina cochlea somatic

data stream

The neocortex learns a model from sensor data - predictions - anomalies - actions

The neocortex learns a sensory-motor model of the world

Principles of Neocortical Function

1) On-line learning from streaming data

retina cochlea somatic

data stream

Principles of Neocortical Function

1) On-line learning from streaming data 2) Hierarchy of memory regions - regions are nearly identical retina cochlea somatic

data stream

Principles of Neocortical Function

1) On-line learning from streaming data 2) Hierarchy of memory regions retina cochlea somatic

data stream

3) Sequence memory - inference - motor

Principles of Neocortical Function

1) On-line learning from streaming data 2) Hierarchy of memory regions 3) Sequence memory

retina cochlea somatic

data stream

4) Sparse Distributed Representations

Principles of Neocortical Function

1) On-line learning from streaming data 2) Hierarchy of memory regions 3) Sequence memory

retina cochlea

data stream

4) Sparse Distributed Representations 5) All regions are sensory and motor

somatic

Motor

Principles of Neocortical Function

1) On-line learning from streaming data 2) Hierarchy of memory regions retina cochlea somatic

data stream

xx xxx xx xx

x

3) Sequence memory

x xx

4) Sparse Distributed Representations 5) All regions are sensory and motor 6) Attention

Principles of Neocortical Function

1) On-line learning from streaming data 2) Hierarchy of memory regions 3) Sequence memory

retina data stream

cochlea

4) Sparse Distributed Representations 5) All regions are sensory and motor

somatic

6) Attention

These six principles are necessary and sufficient for biological and machine intelligence. -

All mammals from mouse to human have them We can build machines like this

Dense Representations • • •

Few bits (8 to 128) All combinations of 1’s and 0’s Example: 8 bit ASCII 01101101 = m

• •

Individual bits have no inherent meaning Representation is arbitrary

Sparse Distributed Representations (SDRs) • • •

Many bits (thousands) Few 1’s mostly 0’s Example: 2,000 bits, 2% active

• •

Each bit has semantic meaning (learned) Representation is semantic

01000000000000000001000000000000000000000000000000000010000…………01000

SDR Properties

1) Similarity: shared bits = semantic similarity

2) Store and Compare: store indices of active bits

subsampling is OK

3) Union membership:

Indices 1 2 3 4 5 | 40

Indices 1 2 | 10

1) 2) 3)

2% ….

10)

Union Is this SDR a member?

20%

Sequence Memory (for inference and motor)

Coincidence detectors

How does a layer of neurons learn sequences?

Each cell is one bit in our Sparse Distributed Representation

SDRs are formed via a local competition between cells. All processes are local across large sheets of cells.

SDR (time =1)

SDR (time =2)

Cells connect to sample of previously active cells to predict their own future activity.

Multiple Predictions Can Occur at Once.

This is a 1st order memory. We need a high order memory.

High order sequences are enabled with multiple cells per column.

High Order Sequence Memory

0 1 0 0 0 0 0 0 1 0 0 0 0 0 0 0 0 0 1 0 0 1 0 0 0 0…………0 1 0 0 0

0 1 0 0 0 0 0 0 1 0 0 0 0 0 0 0 0 0 1 0 0 1 0 0 0 0…………0 1 0 0 0

40 active columns, 10 cells per column = 1040 ways to represent the same input in different contexts A-B-C-D-E X-B’-C’-D’-Y

High Order Sequence Memory

Distributed sequence memory High order, high capacity Noise and fault tolerant Multiple simultaneous predictions Semantic generalization

Online learning • •

Learn continuously, no batch processing If pattern repeats, reinforce, otherwise forget it

Learning is the growth of new synapses.

Connection permanence

unconnected connected

0

0.2

Connection strength is binary Connection permanence is a scalar Training changes permanence

1

“Cortical Learning Algorithm” (CLA)

Not your typical computer memory! A building block for - neocortex - machine intelligence

2 mm

2 mm

Cortical Region

sequence memory sequence memory sequence memory sequence memory

CLA CLA CLA CLA

Feedforward inference Feedforward inference Motor output Feedback / attention

Evidence suggests each layer is implementing a CLA variant

What Is Next? Three Current Directions

1) Commercialization - GROK: Predictive analytics using CLA

- Commercial value accelerates interest and investment 2) Open Source Project - NuPIC: CLA open source software and community - Improve algorithms, develop applications 3) Custom CLA Hardware - Needed for scaling research and commercial applications - IBM, Seagate, Sandia Labs, DARPA

GROK: Predictive Analytics Using CLA

Field 1 Field 2 Field 3 Field N

Field 1 Field 2 Field 3 Field N

Field 1 Field 2 Field 3 Field N

numbers categories text date time

encoder encoder

SDRs

encoder encoder

Encoders Convert native data type to SDRs

Sequence Memory 2,000 cortical columns 60,000 neurons

Predictions Anomalies

Actions

- variable order - online learning

CLA Learns spatial/temporal patterns Outputs - predictions anomalies

GROK example: Factory Energy Usage

Customer need

At midnight, make 24 hourly predictions

GROK Predictions and Actuals

GROK example: Predicting Server Demand

Grok used to predict server demand

Actual Predicted

Approximately 15% reduction in AWS cost

Date

Server demand, Actual vs. Predicted

GROK example: Detecting Anomalous Behavior

Grok builds model of data, detects changes in predictability.

Gear bearing temperature & Grok Anomaly Score

GROK going to market for anomaly detection in I.T. 2014

2) Open Source Project NuPIC: www.Numenta.org - CLA source code (single tree), GPLv3 - Papers, videos, docs Community - 200+ mail list subscribers, growing - 20+ messages per day - full time manager, Matt Taylor

What you can do - Get educated - New applications for CLA - Extend CLA: robotics, language, vision - Tools, documentation 2nd Hackathon November 2,3 in San Francisco - Natural language processing using SDRs - Sensory-motor integration discussion - 2014 hackathon Ireland?

3) Custom CLA Hardware HW companies looking “Beyond von Neumann” - Distributed memory - Fault tolerant - Hierarchical New HW Architectures Needed - Speed (research) - Cost, power, embedded (commercical) IBM

DARPA

- Almaden Research Labs - Joint research agreement - New Program called “Cortical Processor” - HTM (Hierarchical Temporal Memory) - CLA is prototype primitive

Seagate Sandia Labs

Future of Machine Intelligence

Future of Machine Intelligence Definite -

Faster, Bigger Super senses Fluid robotics Distributed hierarchy

Maybe -

Humanoid robots Computer/Brain interfaces for all

Not -

Uploaded brains Evil robots

Why Create Intelligent Machines?

Live better

Learn more

Thank You