Incremental learning by message passing in hierarchical temporal memory.

Area of Science:

• Computational Neuroscience
• Machine Learning
• Artificial Intelligence

Background:

• Hierarchical Temporal Memory (HTM) is a biologically inspired unsupervised learning framework.
• Current HTM models face limitations with incremental learning due to frozen network structures post-training.

Purpose of the Study:

• To develop a novel technique for incremental supervised learning in HTM.
• To enable effective error minimization and backpropagation within the HTM framework.

Main Methods:

• Developed a gradient descent error minimization technique for HTM.
• Implemented error backpropagation using native HTM message passing and belief propagation.
• Utilized a two-stage training approach: unsupervised pretraining followed by supervised refinement.

Main Results:

• Demonstrated the natural and elegant implementation of backpropagation via HTM message passing.
• Experimental results show the two-stage training approach is both accurate and efficient.
• The findings align with recent advancements in other deep learning architectures.

Conclusions:

• The novel supervised learning technique significantly enhances HTM's incremental learning capabilities.
• The proposed method offers an effective and efficient approach for training HTM networks.
• This work contributes to advancing biologically inspired AI and pattern recognition.