Beijing Normal University December, 2015 Yuwei Cui [email protected] Why neurons have thousands of synapses? A model of sequence memory in the brain Collaborators: Jeff Hawkins (PI) Subutai Ahmad Chetan Surpur
Welcome message from author
This document is posted to help you gain knowledge. Please leave a comment to let me know what you think about it! Share it to your friends and learn new things together.
Transcript
Beijing Normal UniversityDecember, 2015
Yuwei [email protected]
Why neurons have thousands of synapses?
A model of sequence memory in the brain
Collaborators: Jeff Hawkins (PI) Subutai Ahmad Chetan Surpur
History
2005 – 2009 HTM theory First generation algorithms Hierarchy and vision problems Vision Toolkit
2002
2004
2009 – 2012 Cortical Learning
Algorithms SDRs, sequence memory,
continuous learning Applications exploration
2013 – 2015 Continued HTM
development NuPIC open source project Grok for anomaly detection
20052014 – ?? Sensorimotor Goal directed
behavior Sequence
classification
Numenta ResearchHTM theoryHTM algorithms
NuPIC
Open source community
Technology Validation and Development
Streaming AnalyticsNatural LanguageSensorimotor Inference
Numenta’s Approach
*HTM = Hierarchical Temporal Memory
NeuroscienceExperimentalResearch
1) Reverse Engineer the Neocortex
- information and biological theory- making good progress
2) Create Technology for Machine Intelligence based on neocortical principles
- not whole-brain simulation, not human-like- new senses, new embodiments, faster , larger
Numenta’s Goals
Mission: Be the leader in the coming era of machine intelligence
What Does the Neocortex Do?
Sensory stream
retina
cochlea
somatic
The neocortex learns a model of the world, primarily through behavior.
Sensory arrays
Motor streamThe model is time-based and predictive.
Top three neocortical principles1) Memory-prediction2) Continuous learning3) Sensory-motor integration
Cortical Architecture
Hierarchy
Cellular layersMini-columns
Neurons: 5-10K synapses
Active dendritesLearning = new synapses
Remarkably uniform - anatomically - functionally
2.5 mm
Sheet of cells
2/34
65
The Neuron
Σ
ANN neuron
Few synapses
Sum input x weights
Learn by modifying weights of synapses
HTM neuron
Thousands of synapses
Active dendrites: Cell recognizes 100’s of unique patterns
Learn by modeling growth of new synapses
Biological neuron
Thousands of synapses
Active dendrites: Cell recognizes 100’s of unique patterns
Learn by growing new synapses
Feedback
Local
FeedforwardLinearGenerate spikes
Non-linear
8-20 coactive synapses lead to dendritic NMDA spikes
Weakly depolarize soma
Hawkins & Ahmad, arXiv 2015
High Order Sequences
Two sequences: A-B-C-DX-B-C-Y
Hawkins & Ahmad, arXiv 2015
B input C input D’ AND Y” predicted
Multiple simultaneous predictions
C’ AND C” predicted
C’ predicted
Prediction of next input
A input B’ predicted B input
Sequence Prediction
Two sequences: A-B-C-DX-B-C-Y
Hawkins & Ahmad, arXiv 2015
1) On-line learning
2) High-order representationsFor example: sequences “ABCD” vs. “XBCY”
3) Multiple simultaneous predictionsFor example: “BC” predicts both “D” and “Y”
4) Fully local and unsupervised learning rules
5) Extremely robustTolerant to >40% noise and faults
6) High capacity
HTM Sequence Memory : Computational Properties
Extensively tested, deployed in commercial applicationsFull source code and documentation available: numenta.org & github.com/numenta Paper in progress, arXiv version available: (Hawkins & Ahmad, 2015; Cui et al, 2015)
Performance On Real-World Streaming Data Sources
ARIMA (statistical method)
RecurrentNeural network(LSTM)
HTM
NYC Taxi demand
Cui et al, arXiv 2015
On-line learning
HTM
Cui et al, arXiv 2015
Ability to Make Multiple Predictions
Sequence Noise Sequence Noise ……
Test Prediction Accuracy
Cui et al, arXiv 2015
Ability to Make Multiple Predictions
Cui et al, arXiv 2015
Fault Tolerance
Datacenterserver anomalies
Rogue human behavior
Geospatial tracking
Stock anomalies
Applications Using HTM High-Order Inference
Social media streams (Twitter)
HTM High OrderSequence Memory
Encoder
SDRData PredictionsAnomalies
Summary- Experimental findings from Neuroscience can lead to improved
learning algorithms - Used properties of active dendrites, Hebbian-style plasticity and minicolumns
- Creating biologically inspired algorithms that really work leads to deeper understanding of cortical principles and numerous testable predictions
Research Roadmap- Understand functional properties of laminar microcircuit and
thalamocortical inputs- Model multiple regions and hierarchy- More biophysically accurate neuron models (e.g. spiking models)
Numenta Licensees
Cortical.ioNatural language processing using HTM principleswww.Cortical.io
GrokStreamIT monitoring using HTMwww.GrokStream.com
Numenta Research Partnerships
IBM ResearchCreating complete technology stack for HTM systemsLead: Dr. Winfried Wilcke
DARPAHTM-based “Cortical Processor”Lead: Dr. Dan Hammerstrom
University of HeidelbergPorted HTM sequence memory to HICANN neuromorphic chipLead: Dr. Karlheinz Meier
University of BerlinTesting biological predictions of HTM theoryLead: Dr. Matthew Larkum
1) Sparser activations during a predictable sensory stream.
2) Unanticipated inputs leads to a burst of activity correlated vertically within mini-columns.
3) Neighboring mini-columns will not be correlated.
4) Predicted cells need fast inhibition to inhibit nearby cells within mini-column.
5) For predictable stimuli, dendritic NMDA spikes will be much more frequent than somatic action potentials.
6) Localized synaptic plasticity for dendritic segments that have spiked followed a short time later by a back action potential.
7) The existence of sub-threshold LTP (in the absence of NMDA spikes) in dendritic segments if a cluster of synapses become active followed by a bAP.
8) The existence of localized weak LTD when an NMDA spike is not followed by an action potential.
Testable Predictions
(Vinje & Gallant, 2002)
(Ecker et al, 2010; Smith & Häusser, 2010)
(Smith et al, 2013)
(Losonczy et al, 2008)
Summary- Experimental findings from Neuroscience can lead to improved
learning algorithms - Used properties of active dendrites, Hebbian-style plasticity and minicolumns
- Creating biologically inspired algorithms that really work leads to deeper understanding of cortical principles and numerous testable predictions
Research Roadmap- Understand functional properties of laminar microcircuit and
thalamocortical inputs- Model multiple regions and hierarchy- More biophysically accurate neuron models (e.g. spiking models)
Comparison With Common Sequence Memory Algorithms
Fault Tolerance
Branco, T., & Häusser, M. (2011). Synaptic integration gradients in single cortical pyramidal cell dendrites. Neuron, 69(5), 885–92.
NMDA Dendritic Spike
Local
Active Dendrites - Highlights
Feedforward
Feedback
Experimental DataSynapses on distal segments have a non-linear effect.
8 to 20 coactive synapses on a distal dendrite branch will cause an NMDA dendritic spike. (This is a small fraction of spines on the branch.)
Synapse activity must be spatially and temporally localized
NMDA spike will depolarize soma but not cause action potential.
85% of excitatory synapses on distal dendrites.
(Branco & Häusser, 2011; Schiller et al, 2000; Losonczy, 2006; Antic et al, 2010; Major et al, 2013; Spruston, 2008; Milojkovic et al, 2005, etc.)
Related Documents
