On-line Learning From Streaming Data
ACM CIKM
October 31, 2013
Jeff Hawkins
jhawkins@GrokSolutions.com
Industrial Research Track
1) Discover operating principles of neocortex
Anatomy,
Physiology
Theoretical
principles
Software
2) Build systems based on these principles
Cortical algorithms
Anomaly detection in high velocity data
The neocortex is a memory system.
retina
cochlea
somatic
data stream
The neocortex learns a model
from sensory 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
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
somatic
4) Sparse Distributed Representations
5) All regions are sensory and motor
6) Attention
These six principles are necessary and sufficient
for biological and machine intelligence.
- All mammals from mouse to human have them
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 assigned by programmer
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
A Few SDR Properties
1) Similarity:
shared bits = semantic similarity
2) Store and Compare:
store indices of active bits
subsampling is OK
Indices
1
2
3
4
5
|
40
Indices
1
2
|
10
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.
SDR (time =1)
SDR (time =2)
Cell forms connections to subsample of previously active cells.
Predicts its own future activity.
Multiple Predictions Can Occur at Once
With one cell per column, 1st order memory
We need a high order memory
High Order Sequence Memory
Enabled by Columns of Cells
Cortical Learning Algorithm (CLA)
Distributed sequence memory
High order
High capacity
Multiple simultaneous predictions
Semantic generalization
Three Current Directions
1) NuPIC Open Source Project
NuPIC Open Source Project
www.Numenta.org
Single source tree (used by GROK)
GPLv3
Steady community growth
–
–
–
67 contributors (+26 since July)
245 mailing list subscribers
1621 total messages
eBook from community member
OS community joining Kaggle Competitions
Fall Hackathon: 70 attendees
Three Current Directions
1) NuPIC Open Source Project
2) Custom CLA Hardware
- Needed for scaling research and commercial applications
- DARPA “Cortical Processor”
- IBM, Seagate, Sandia Labs
3) Commercialization
Data: Past and Future
Past
1. Store data
Future
2. Look at data
3. Build models
Problem:
- Doesn’t scale with velocity and #models
Solution:
- Automated model creation
- Continuous learning
- Temporal inference
Stream data
Automated model creation
Continuous learning
Temporal inference
Predictions
Anomalies
Actions
Anomaly Detection Using Predictive Cortical Models
Cortical Memory
Metric 1
Encoder
SDR
Prediction
Point anomaly score
Time average
Distribution of averages
Metric anomaly score
System
Anomaly
Score
.
.
.
Cortical Memory
Metric N
Encoder
SDR
Prediction
Point anomaly score
Time average
Distribution of averages
Metric anomaly score
Anomaly
score
Metric value
Anomaly
score
Metric value
Largely predictable
Largely unpredictable
Grok for IT Monitoring
Breakthrough Science
for Anomaly Detection
 Detects problems thresholds
miss
 Continuous learning
 Automated model building
 State-of-the art neocortical
model
Reinventing UX for
IT Monitoring
 Smartphone-centric
 Ranks anomalous instances
 Rapid drill down
 Continuously updated
 User-controlled notifications
In private beta for Amazon AWS cloud users
grokbeta@GrokSolutions.com
Extensible Architecture
 Custom metrics for any
application/server
 Web interface and mobile client
source code available under no-cost
license
 Engine API to be published
 NuPIC open source community