Continuous Online Sequence Learning with an Unsupervised Neural Network Model.

Area of Science:

• Neuroscience
• Machine Learning
• Computational Neuroscience

Background:

• Recognizing and predicting temporal sequences is crucial for survival.
• Hierarchical Temporal Memory (HTM) is a theoretical framework for cortical sequence learning.
• HTM models properties of cortical neurons for sequence processing.

Purpose of the Study:

• Analyze properties of HTM sequence memory.
• Apply HTM to sequence learning and prediction with streaming data.
• Evaluate HTM's performance against other sequence learning algorithms.

Main Methods:

• Utilized an unsupervised Hebbian-like learning rule for continuous learning.
• Employed sparse temporal codes for robust handling of branching sequences.
• Compared HTM with statistical, feedforward, and recurrent neural network models.

Main Results:

• HTM continuously learns numerous variable-order temporal sequences.
• The model maintains multiple predictions for branching sequences.
• HTM achieved accuracy comparable to state-of-the-art algorithms on diverse datasets.

Conclusions:

• HTM sequence memory offers continuous online learning and robustness to noise.
• The model effectively handles multiple predictions and high-order statistics in sequences.
• HTM provides insights into brain function and is applicable to real-world streaming data problems.