The HTM Spatial Pooler-A Neocortical Algorithm for Online Sparse Distributed Coding.

Area of Science:

• Computational Neuroscience
• Machine Learning
• Artificial Intelligence

Background:

• The neocortex's computational principles are modeled by Hierarchical Temporal Memory (HTM).
• The HTM spatial pooler (SP) is a key component for learning feedforward connections and creating efficient input representations.

Purpose of the Study:

• Analyze the HTM spatial pooler (SP) and its properties.
• Develop metrics to quantify SP performance.
• Demonstrate the SP's value in a complete HTM system.

Main Methods:

• The SP converts binary input patterns into sparse distributed representations (SDRs).
• It employs competitive Hebbian learning and homeostatic excitability control.
• Metrics were developed to quantify properties like adaptation, noise robustness, cell efficiency, and cell death robustness.

Main Results:

• The SP exhibits fast adaptation to changing input statistics.
• It shows improved noise robustness through learning.
• The SP demonstrates efficient cell utilization and robustness to cell death.

Conclusions:

• The HTM spatial pooler is a neurally inspired algorithm for online learning of sparse representations from noisy data.
• The developed metrics effectively quantify the SP's key properties.
• The SP proves valuable in real-world HTM applications.