Sequential activity in asymmetrically coupled winner-take-all circuits.

Area of Science:

• Computational Neuroscience
• Neural Networks
• Machine Learning

Background:

• Recurrent neural network models for sequence learning often lack neurophysiological realism and robustness.
• Existing models struggle with noise and require fine-tuning, limiting their biological plausibility.

Purpose of the Study:

• To propose a novel, neurobiologically plausible model for sequence learning and generation.
• To develop a network architecture consistent with mammalian cortical connectivity and realistic neuronal dynamics.
• To achieve noise-robust sequential activity that can be controlled and modulated by external inputs.

Main Methods:

• Utilized multiple asymmetrically coupled winner-take-all (WTA) circuits.
• Incorporated realistic neuronal and synaptic dynamics, including conductance-based synapses and spike-timing dependent plasticity (STDP).
• Analyzed a rate-based approximation and validated with numerical simulations of spiking neurons.

Main Results:

• The proposed network generates robust sequential activity patterns (activity bumps) that propagate through WTA circuits.
• The model exhibits noise-robustness and allows for halting, resuming, and external modulation of sequences.
• Demonstrated learning and reproduction of arbitrary sequential patterns using plastic synapses and STDP.

Conclusions:

• The novel WTA-based network offers a neurobiologically plausible mechanism for sequence learning and generation.
• The model's ability to control and learn sequences provides a foundation for state-dependent perception-action loops.
• The findings support the use of realistic neural dynamics and plasticity for robust sequence processing.