Liquid state machine with dendritically enhanced readout for low-power, neuromorphic VLSI implementations.

Area of Science:

• Computational Neuroscience
• Neuromorphic Engineering
• Machine Learning

Background:

• Liquid State Machines (LSMs) are popular for reservoir computing.
• Current readout stages, like parallel perceptrons, offer high performance but require significant synaptic resources.
• Biological neurons exhibit nonlinear dendritic properties that inspire novel computational architectures.

Purpose of the Study:

• To develop a new neuro-inspired, hardware-friendly readout stage for LSMs.
• To improve performance and reduce synaptic resource requirements compared to existing methods.
• To leverage biological neuron principles for enhanced computational efficiency.

Main Methods:

• Proposed a readout architecture with multi-dendrite neurons featuring lumped nonlinearity (two-compartment model).
• Implemented a learning algorithm involving network rewiring (NRW) to optimize synaptic connections on dendritic branches.
• Utilized binary synapses and explored address event representation (AER) protocols for hardware implementation.

Main Results:

• Achieved up to 3.3x less error in spike train classification and 2.4x less error in input rate approximation compared to single perceptrons with analog weights, using binary synapses.
• Demonstrated superior performance over 60 parallel perceptrons, even with significantly larger synapses.
• Showcased robustness against statistical variations due to binary synapses.

Conclusions:

• The proposed dendritically enhanced readout stage offers a more efficient and performant alternative for reservoir computing.
• The architecture is attractive for VLSI and neuromorphic implementations due to reduced resource needs and compatibility with AER protocols.
• Network rewiring inspired by structural plasticity enhances learning and adaptability in neuromorphic systems.