computing like the brain: the path to machine intelligence · 17-09-2013 · [email protected]...
TRANSCRIPT
![Page 1: Computing Like the Brain: The path to machine intelligence · 17-09-2013 · jhawkins@numenta.com Computing Like the Brain: The path to machine intelligence . 1) Discover operating](https://reader033.vdocuments.mx/reader033/viewer/2022060223/5f07f0517e708231d41f841f/html5/thumbnails/1.jpg)
NASA September 17, 2013
Jeff Hawkins [email protected]
Computing Like the Brain: The path to machine intelligence
![Page 2: Computing Like the Brain: The path to machine intelligence · 17-09-2013 · jhawkins@numenta.com Computing Like the Brain: The path to machine intelligence . 1) Discover operating](https://reader033.vdocuments.mx/reader033/viewer/2022060223/5f07f0517e708231d41f841f/html5/thumbnails/2.jpg)
1) Discover operating principles of neocortex 2) Build systems based on these principles
![Page 3: Computing Like the Brain: The path to machine intelligence · 17-09-2013 · jhawkins@numenta.com Computing Like the Brain: The path to machine intelligence . 1) Discover operating](https://reader033.vdocuments.mx/reader033/viewer/2022060223/5f07f0517e708231d41f841f/html5/thumbnails/3.jpg)
Alan Turing
“Computers are universal machines” 1935+ “Human behavior as test for machine intelligence” 1950
Pros: - Good solutions Cons: - Task specific
- Limited or no learning
Artificial Intelligence - no neuroscience
• MIT AI Lab • 5th Generation Computing Project • DARPA Strategic Computing Initiative • DARPA Grand Challenge
Major AI Initiatives
AI Projects
• ACT-R • Asimo • CoJACK • Cyc • Deep Blue • Global Workspace Theory • Mycin • SHRDLU • Soar • Watson - Many more -
![Page 4: Computing Like the Brain: The path to machine intelligence · 17-09-2013 · jhawkins@numenta.com Computing Like the Brain: The path to machine intelligence . 1) Discover operating](https://reader033.vdocuments.mx/reader033/viewer/2022060223/5f07f0517e708231d41f841f/html5/thumbnails/4.jpg)
Artificial Neural Networks – minimal neuroscience
Warren McCulloch Walter Pitts
“Neurons as logic gates” 1943 Proposed first artificial neural network
ANN techniques
Pros: - Good classifiers - Learning systems
Cons: - Limited capabilities
- Not brain like
• Back propagation • Boltzman machines • Hopfield networks • Kohonen networks • Parallel Distributed Processing • Machine learning • Deep Learning
![Page 5: Computing Like the Brain: The path to machine intelligence · 17-09-2013 · jhawkins@numenta.com Computing Like the Brain: The path to machine intelligence . 1) Discover operating](https://reader033.vdocuments.mx/reader033/viewer/2022060223/5f07f0517e708231d41f841f/html5/thumbnails/5.jpg)
Whole Brain Simulator – maximal neuroscience
The Human Brain Project
No theory No attempt at Machine Intelligence
Blue Brain simulation
![Page 6: Computing Like the Brain: The path to machine intelligence · 17-09-2013 · jhawkins@numenta.com Computing Like the Brain: The path to machine intelligence . 1) Discover operating](https://reader033.vdocuments.mx/reader033/viewer/2022060223/5f07f0517e708231d41f841f/html5/thumbnails/6.jpg)
1) Discover operating principles of neocortex 2) Build systems based on these principles
Anatomy, Physiology
Theoretical principles
Software Silicon
![Page 7: Computing Like the Brain: The path to machine intelligence · 17-09-2013 · jhawkins@numenta.com Computing Like the Brain: The path to machine intelligence . 1) Discover operating](https://reader033.vdocuments.mx/reader033/viewer/2022060223/5f07f0517e708231d41f841f/html5/thumbnails/7.jpg)
The neocortex is a memory system.
data stream retina
cochlea
somatic
The neocortex learns a model from sensory data - predictions - anomalies - actions
The neocortex learns a sensory-motor model of the world
![Page 8: Computing Like the Brain: The path to machine intelligence · 17-09-2013 · jhawkins@numenta.com Computing Like the Brain: The path to machine intelligence . 1) Discover operating](https://reader033.vdocuments.mx/reader033/viewer/2022060223/5f07f0517e708231d41f841f/html5/thumbnails/8.jpg)
Principles of Neocortical Function
retina
cochlea
somatic
1) On-line learning from streaming data
data stream
![Page 9: Computing Like the Brain: The path to machine intelligence · 17-09-2013 · jhawkins@numenta.com Computing Like the Brain: The path to machine intelligence . 1) Discover operating](https://reader033.vdocuments.mx/reader033/viewer/2022060223/5f07f0517e708231d41f841f/html5/thumbnails/9.jpg)
Principles of Neocortical Function
2) Hierarchy of memory regions
retina
cochlea
somatic
1) On-line learning from streaming data
data stream
![Page 10: Computing Like the Brain: The path to machine intelligence · 17-09-2013 · jhawkins@numenta.com Computing Like the Brain: The path to machine intelligence . 1) Discover operating](https://reader033.vdocuments.mx/reader033/viewer/2022060223/5f07f0517e708231d41f841f/html5/thumbnails/10.jpg)
Principles of Neocortical Function
2) Hierarchy of memory regions
retina
cochlea
somatic
3) Sequence memory - inference - motor
data stream
1) On-line learning from streaming data
![Page 11: Computing Like the Brain: The path to machine intelligence · 17-09-2013 · jhawkins@numenta.com Computing Like the Brain: The path to machine intelligence . 1) Discover operating](https://reader033.vdocuments.mx/reader033/viewer/2022060223/5f07f0517e708231d41f841f/html5/thumbnails/11.jpg)
Principles of Neocortical Function
4) Sparse Distributed Representations
2) Hierarchy of memory regions
retina
cochlea
somatic
3) Sequence memory data stream
1) On-line learning from streaming data
![Page 12: Computing Like the Brain: The path to machine intelligence · 17-09-2013 · jhawkins@numenta.com Computing Like the Brain: The path to machine intelligence . 1) Discover operating](https://reader033.vdocuments.mx/reader033/viewer/2022060223/5f07f0517e708231d41f841f/html5/thumbnails/12.jpg)
Principles of Neocortical Function
retina
cochlea
somatic
data stream
2) Hierarchy of memory regions
3) Sequence memory
5) All regions are sensory and motor
4) Sparse Distributed Representations
Motor
1) On-line learning from streaming data
![Page 13: Computing Like the Brain: The path to machine intelligence · 17-09-2013 · jhawkins@numenta.com Computing Like the Brain: The path to machine intelligence . 1) Discover operating](https://reader033.vdocuments.mx/reader033/viewer/2022060223/5f07f0517e708231d41f841f/html5/thumbnails/13.jpg)
Principles of Neocortical Function
retina
cochlea
somatic
data stream
x x x x x x
x x x x x x x
2) Hierarchy of memory regions
3) Sequence memory
5) All regions are sensory and motor
6) Attention
4) Sparse Distributed Representations
1) On-line learning from streaming data
![Page 14: Computing Like the Brain: The path to machine intelligence · 17-09-2013 · jhawkins@numenta.com Computing Like the Brain: The path to machine intelligence . 1) Discover operating](https://reader033.vdocuments.mx/reader033/viewer/2022060223/5f07f0517e708231d41f841f/html5/thumbnails/14.jpg)
Principles of Neocortical Function
4) Sparse Distributed Representations
2) Hierarchy of memory regions
retina
cochlea
somatic
3) Sequence memory
5) All regions are sensory and motor
6) Attention
data stream
1) On-line learning from streaming data
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
![Page 15: Computing Like the Brain: The path to machine intelligence · 17-09-2013 · jhawkins@numenta.com Computing Like the Brain: The path to machine intelligence . 1) Discover operating](https://reader033.vdocuments.mx/reader033/viewer/2022060223/5f07f0517e708231d41f841f/html5/thumbnails/15.jpg)
Sparse Distributed Representations (SDRs)
• Many bits (thousands) • Few 1’s mostly 0’s • Example: 2,000 bits, 2% active
• Each bit has semantic meaning • Meaning of each bit is learned, not assigned
01000000000000000001000000000000000000000000000000000010000…………01000
Dense Representations
• Few bits (8 to 128) • All combinations of 1’s and 0’s • Example: 8 bit ASCII
• Individual bits have no inherent meaning • Representation is assigned by programmer
01101101 = m
![Page 16: Computing Like the Brain: The path to machine intelligence · 17-09-2013 · jhawkins@numenta.com Computing Like the Brain: The path to machine intelligence . 1) Discover operating](https://reader033.vdocuments.mx/reader033/viewer/2022060223/5f07f0517e708231d41f841f/html5/thumbnails/16.jpg)
SDR Properties
1) Similarity: shared bits = semantic similarity
subsampling is OK
3) Union membership:
Indices 1 2 | 10
Is this SDR a member?
2) Store and Compare: store indices of active bits
Indices 1 2 3 4 5 | 40
1) 2) 3)
…. 10)
2%
20%
Union
![Page 17: Computing Like the Brain: The path to machine intelligence · 17-09-2013 · jhawkins@numenta.com Computing Like the Brain: The path to machine intelligence . 1) Discover operating](https://reader033.vdocuments.mx/reader033/viewer/2022060223/5f07f0517e708231d41f841f/html5/thumbnails/17.jpg)
Coincidence detectors
How does a layer of neurons learn sequences?
Sequence Memory (for inference and motor)
![Page 18: Computing Like the Brain: The path to machine intelligence · 17-09-2013 · jhawkins@numenta.com Computing Like the Brain: The path to machine intelligence . 1) Discover operating](https://reader033.vdocuments.mx/reader033/viewer/2022060223/5f07f0517e708231d41f841f/html5/thumbnails/18.jpg)
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.
![Page 19: Computing Like the Brain: The path to machine intelligence · 17-09-2013 · jhawkins@numenta.com Computing Like the Brain: The path to machine intelligence . 1) Discover operating](https://reader033.vdocuments.mx/reader033/viewer/2022060223/5f07f0517e708231d41f841f/html5/thumbnails/19.jpg)
SDR (time =1)
![Page 20: Computing Like the Brain: The path to machine intelligence · 17-09-2013 · jhawkins@numenta.com Computing Like the Brain: The path to machine intelligence . 1) Discover operating](https://reader033.vdocuments.mx/reader033/viewer/2022060223/5f07f0517e708231d41f841f/html5/thumbnails/20.jpg)
SDR (time =2)
![Page 21: Computing Like the Brain: The path to machine intelligence · 17-09-2013 · jhawkins@numenta.com Computing Like the Brain: The path to machine intelligence . 1) Discover operating](https://reader033.vdocuments.mx/reader033/viewer/2022060223/5f07f0517e708231d41f841f/html5/thumbnails/21.jpg)
Cells form connections to subsample of previously active cells. Predicts its own future activity.
![Page 22: Computing Like the Brain: The path to machine intelligence · 17-09-2013 · jhawkins@numenta.com Computing Like the Brain: The path to machine intelligence . 1) Discover operating](https://reader033.vdocuments.mx/reader033/viewer/2022060223/5f07f0517e708231d41f841f/html5/thumbnails/22.jpg)
Multiple Predictions Can Occur at Once
With one cell per column, 1st order memory We need a high order memory
![Page 23: Computing Like the Brain: The path to machine intelligence · 17-09-2013 · jhawkins@numenta.com Computing Like the Brain: The path to machine intelligence . 1) Discover operating](https://reader033.vdocuments.mx/reader033/viewer/2022060223/5f07f0517e708231d41f841f/html5/thumbnails/23.jpg)
High order sequences are enabled with multiple cells per column.
![Page 24: Computing Like the Brain: The path to machine intelligence · 17-09-2013 · jhawkins@numenta.com Computing Like the Brain: The path to machine intelligence . 1) Discover operating](https://reader033.vdocuments.mx/reader033/viewer/2022060223/5f07f0517e708231d41f841f/html5/thumbnails/24.jpg)
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
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
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
![Page 25: Computing Like the Brain: The path to machine intelligence · 17-09-2013 · jhawkins@numenta.com Computing Like the Brain: The path to machine intelligence . 1) Discover operating](https://reader033.vdocuments.mx/reader033/viewer/2022060223/5f07f0517e708231d41f841f/html5/thumbnails/25.jpg)
High Order Sequence Memory
Distributed sequence memory Works across large areas High order, high capacity Multiple simultaneous predictions Semantic generalization
![Page 26: Computing Like the Brain: The path to machine intelligence · 17-09-2013 · jhawkins@numenta.com Computing Like the Brain: The path to machine intelligence . 1) Discover operating](https://reader033.vdocuments.mx/reader033/viewer/2022060223/5f07f0517e708231d41f841f/html5/thumbnails/26.jpg)
Online learning
• Learn continuously, no batch processing • If pattern repeats, reinforce, otherwise forget it
Connection strength is binary Connection permanence is a scalar Training changes permanence
Learning is the growth of new synapses.
1 0
connected unconnected
Connection permanence 0.2
![Page 27: Computing Like the Brain: The path to machine intelligence · 17-09-2013 · jhawkins@numenta.com Computing Like the Brain: The path to machine intelligence . 1) Discover operating](https://reader033.vdocuments.mx/reader033/viewer/2022060223/5f07f0517e708231d41f841f/html5/thumbnails/27.jpg)
Feedforward inference Feedforward inference
Motor output Feedback / a4en5on
2 mm
2 mm
Cortical Region
![Page 28: Computing Like the Brain: The path to machine intelligence · 17-09-2013 · jhawkins@numenta.com Computing Like the Brain: The path to machine intelligence . 1) Discover operating](https://reader033.vdocuments.mx/reader033/viewer/2022060223/5f07f0517e708231d41f841f/html5/thumbnails/28.jpg)
Feedforward inference Feedforward inference
Motor output Feedback / a4en5on
2 mm
2 mm
Cortical Region
![Page 29: Computing Like the Brain: The path to machine intelligence · 17-09-2013 · jhawkins@numenta.com Computing Like the Brain: The path to machine intelligence . 1) Discover operating](https://reader033.vdocuments.mx/reader033/viewer/2022060223/5f07f0517e708231d41f841f/html5/thumbnails/29.jpg)
Feedforward inference Feedforward inference
Motor output Feedback / a4en5on
2 mm
2 mm
sequence memory sequence memory
sequence memory sequence memory
CLA CLA CLA CLA
Evidence suggests each layer is implementing a CLA variant
Cortical Region
![Page 30: Computing Like the Brain: The path to machine intelligence · 17-09-2013 · jhawkins@numenta.com Computing Like the Brain: The path to machine intelligence . 1) Discover operating](https://reader033.vdocuments.mx/reader033/viewer/2022060223/5f07f0517e708231d41f841f/html5/thumbnails/30.jpg)
1) Open Source Project - NuPIC: CLA open source software and community
- Improve algorithms, develop applications
2) Commercialization
- GROK: Predictive analytics using CLA - Commercial value generates investment $
3) Custom CLA Hardware - Needed for scaling research and commercial applications
- IBM, Seagate, Sandia Labs, DARPA
Three Current Directions
![Page 31: Computing Like the Brain: The path to machine intelligence · 17-09-2013 · jhawkins@numenta.com Computing Like the Brain: The path to machine intelligence . 1) Discover operating](https://reader033.vdocuments.mx/reader033/viewer/2022060223/5f07f0517e708231d41f841f/html5/thumbnails/31.jpg)
www.Numenta.org Single source tree (used by GROK)
GPLv3 license Active community
- 215 mail list subscribers - 20 messages per day - growing - full time manager, Matt Taylor
Next hackathon November 2 & 3 in San Francisco
- NLP using SDRs
- Sensory-motor integration using CLA discussion
NuPIC: CLA Open Source Project
![Page 32: Computing Like the Brain: The path to machine intelligence · 17-09-2013 · jhawkins@numenta.com Computing Like the Brain: The path to machine intelligence . 1) Discover operating](https://reader033.vdocuments.mx/reader033/viewer/2022060223/5f07f0517e708231d41f841f/html5/thumbnails/32.jpg)
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
Sequence Memory 2,000 cortical columns 60,000 neurons - variable order - online learning
encoder
encoder
encoder
encoder
SDRs Predictions Anomalies
GROK: Predictive Analytics Using CLA
Encoders Convert native data type to SDRs
CLA Learns spatial/temporal patterns Outputs - predictions anomalies
Actions
![Page 33: Computing Like the Brain: The path to machine intelligence · 17-09-2013 · jhawkins@numenta.com Computing Like the Brain: The path to machine intelligence . 1) Discover operating](https://reader033.vdocuments.mx/reader033/viewer/2022060223/5f07f0517e708231d41f841f/html5/thumbnails/33.jpg)
GROK example: Factory Energy Usage
![Page 34: Computing Like the Brain: The path to machine intelligence · 17-09-2013 · jhawkins@numenta.com Computing Like the Brain: The path to machine intelligence . 1) Discover operating](https://reader033.vdocuments.mx/reader033/viewer/2022060223/5f07f0517e708231d41f841f/html5/thumbnails/34.jpg)
Customer need
At midnight, make 24 hourly predictions
![Page 35: Computing Like the Brain: The path to machine intelligence · 17-09-2013 · jhawkins@numenta.com Computing Like the Brain: The path to machine intelligence . 1) Discover operating](https://reader033.vdocuments.mx/reader033/viewer/2022060223/5f07f0517e708231d41f841f/html5/thumbnails/35.jpg)
GROK Predictions and Actuals
![Page 36: Computing Like the Brain: The path to machine intelligence · 17-09-2013 · jhawkins@numenta.com Computing Like the Brain: The path to machine intelligence . 1) Discover operating](https://reader033.vdocuments.mx/reader033/viewer/2022060223/5f07f0517e708231d41f841f/html5/thumbnails/36.jpg)
GROK example: Predicting Server Demand
Date
Actual Predicted
Server demand, Actual vs. Predicted
Grok used to predict server demand Approximately 15% reduction in AWS cost
![Page 37: Computing Like the Brain: The path to machine intelligence · 17-09-2013 · jhawkins@numenta.com Computing Like the Brain: The path to machine intelligence . 1) Discover operating](https://reader033.vdocuments.mx/reader033/viewer/2022060223/5f07f0517e708231d41f841f/html5/thumbnails/37.jpg)
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
![Page 38: Computing Like the Brain: The path to machine intelligence · 17-09-2013 · jhawkins@numenta.com Computing Like the Brain: The path to machine intelligence . 1) Discover operating](https://reader033.vdocuments.mx/reader033/viewer/2022060223/5f07f0517e708231d41f841f/html5/thumbnails/38.jpg)
IBM - Almaden Labs
- Joint research agreement - Winfried Wilcke
DARPA - “Cortical Processor”
- “HTM” (Hierarchical Temporal Memory)
- CLA is prototype primitive
- Dan Hammerstrom
Seagate
Sandia Labs
Custom CLA Hardware
![Page 39: Computing Like the Brain: The path to machine intelligence · 17-09-2013 · jhawkins@numenta.com Computing Like the Brain: The path to machine intelligence . 1) Discover operating](https://reader033.vdocuments.mx/reader033/viewer/2022060223/5f07f0517e708231d41f841f/html5/thumbnails/39.jpg)
Future of Machine Intelligence
![Page 40: Computing Like the Brain: The path to machine intelligence · 17-09-2013 · jhawkins@numenta.com Computing Like the Brain: The path to machine intelligence . 1) Discover operating](https://reader033.vdocuments.mx/reader033/viewer/2022060223/5f07f0517e708231d41f841f/html5/thumbnails/40.jpg)
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 - Friendly uses only
![Page 41: Computing Like the Brain: The path to machine intelligence · 17-09-2013 · jhawkins@numenta.com Computing Like the Brain: The path to machine intelligence . 1) Discover operating](https://reader033.vdocuments.mx/reader033/viewer/2022060223/5f07f0517e708231d41f841f/html5/thumbnails/41.jpg)
Why Machine Intelligence?
Thank You
Live better Learn more