Learning compositional sequences with multiple time scales through a hierarchical network of spiking neurons

Area of Science:

• Computational Neuroscience
• Machine Learning
• Systems Neuroscience

Background:

• Sequential behaviors are often compositional, organized hierarchically across multiple timescales.
• Existing neural network models for temporal learning predominantly use serial processing methods.
• Experimental neuroscience provides mounting evidence for hierarchical structures in neural processing.

Purpose of the Study:

• To introduce a novel hierarchical network model of spiking neurons for multi-timescale sequence learning.
• To investigate the model's capacity for independent learning of motifs and syntax.
• To compare the performance of the hierarchical model against serial learning approaches.

Main Methods:

• Development of a spiking neural network model with a hierarchical organization.
• Implementation of biophysically motivated neuron dynamics and local plasticity rules.
• Evaluation of learning speed, relearning flexibility, storage capacity, and robustness to perturbations.

Main Results:

• The hierarchical model successfully learns motifs and syntax independently.
• The model demonstrates faster learning, more flexible relearning, increased storage capacity, and higher robustness compared to serial models.
• Hierarchical learning achieves high motif fidelity by redistributing variability to between-motif timings.

Conclusions:

• Hierarchical organization in neural networks offers significant advantages for multi-timescale sequence learning.
• The proposed model provides a biologically plausible framework for understanding hierarchical temporal processing.
• This approach enhances learning efficiency, adaptability, and memory in artificial neural systems.