Numenta engineer Yuwei Cui walks through how the HTM Spatial Pooler works, explaining why desired properties exist and how they work. Includes lots of graphs of SP online learning performance, discussion of topology and boosting.
2. Spatial Pooler Resources
• Hawkins J, Ahmad S, Dubinsky D. (2011) Cortical learning algorithm and
hierarchical temporal memory. Numenta Whitepaper
http://numenta.org/resources/HTM_CorticalLearningAlgorithms.pdf
• Cui Y, Ahmad S, Hawkins J (under review) The HTM Spatial Pooler: A
Neocortical Algorithm for Online Sparse Distributed Coding.
bioRxiv. DOI: 10.1101/085035
• HTM School: http://numenta.org/htm-school/
• Spatial Pooler in NUPIC:
https://github.com/numenta/nupic/blob/master/src/nupic/research/spatial_pooler.py
3. Background
• How do individual neurons learn to respond to specific input patterns?
• How do populations of neurons represents input features?
retina
cochlea
somatic
data stream
motor control
millions of spike trains!
4. Online sequence learning with HTM
HTM High Order
Sequence Memory
Encoder
SDRData
Predictions
Classification
Classifier
Anomaly
Likelihood Anomalies
Spatial
Pooler
5. HTM Spatial Pooler
HTM High Order
Sequence Memory
Encoder
SDRData
Predictions
Classification
Classifier
Anomaly
Likelihood Anomalies
Spatial
Pooler
Spatial Pooler
Input 1 Input 2 Input 3
SDRs with fixed sparsity
Large number of unlabeled inputs
Variable sparsity
…………
What properties do SP achieve?
10. HTM spatial pooler – learning
Hebbian Learning Rule:
Active (winner) MCs reinforce
their active input connections,
and depress inactive inputs
11. Learning in HTM Spatial Pooler
• Why do we need learning in SP?
If inputs are random, an untrained
random SP will do just as good as any
trained SP
However, real inputs are structured.
Input SDRs occurs with non-equal probabilities.
The actual inputs should be “better” represented
than random inputs after learning
18. Properties of SP – fault tolerance
Experiment 1: lesion of SP MCs
We monitor:
• RF center of SP MCs
• Coverage of input space
• Avg boost factors
• Growth & elimination of synapses
19. Properties of SP – fault tolerance
Experiment 1: lesion of SP MCs
20. Properties of SP – fault tolerance
Experiment 1: lesion of SP MCs
21. Properties of SP – fault tolerance
Experiment 2: lesion of afferent inputs
We monitor:
• RF center of SP MCs
• Coverage of input space
• Avg boost factors
• Growth & elimination of synapses
22. Properties of SP – fault tolerance
Experiment 2: lesion of afferent inputs
23. Properties of SP – fault tolerance
Experiment 2: lesion of afferent inputs
25. Resources
• Hawkins J, Ahmad S, Dubinsky D. Cortical learning algorithm and hierarchical
temporal memory. In: Numenta Whitepaper [Internet]. 2011 pp. 1–68.
Available: http://numenta.org/resources/HTM_CorticalLearningAlgorithms.pdf
• Cui Y, Ahmad S, Hawkins J (under review) The HTM Spatial Pooler: A
Neocortical Algorithm for Online Sparse Distributed Coding. bioRxiv. DOI:
10.1101/085035
The brain is a powerful streaming analytics engine.
Our brain continuously receives vast amounts of information about the external world through peripheral sensors that transform changes in light luminance, sound pressure, and skin deformations into millions of spike trains. Each cortical neuron has to make sense of a flood of time-varying inputs by forming synaptic connections to a subset of the presynaptic neurons.
A central problem in neuroscience is to understand how individual cortical neurons learn to respond to specific input spike patterns, and how a population of neurons collectively represents features of the inputs in a flexible, dynamic, yet robust way.
The current generation of HTM models continuous sequence learning in the brain.
The HTM models a layer of cells. It consists a set of cells that are organized into MCs.
The HTM neuron has three different types of inputs for each neuron.
Because in the cortex cells in the same mcs have the same FF RFs, …
The SP models FF
The SP receives inputs from a set of input neurons. Each MC connects to a subset of the input space.
At any time, a small set of MCs are active in the SP.
These MCs are selected using WTA