biomimetic-snn-rust

biomimetic-snn-rust is a cutting-edge Spiking Neural Network (SNN) simulator developed in pure Rust, designed to be zero-dependency and multi-threaded. This project emerges from a passion for exploring biophysically accurate neuron models and maximizing lock-free parallel execution within CPU environments.

Key Features:

• Architecture Optimized for Performance: Employing a Data-Oriented Design (DoD) layout, this simulator leverages the Rust standard library for minimal overhead in execution.
• Neuron Dynamics: Utilizes the Leaky Integrate-and-Fire (LIF) model, complete with adaptive thresholding and both absolute and relative refractory states, enabling realistic neuron behavior.
• Synaptic Plasticity Mechanisms:
  • Spike-Timing-Dependent Plasticity (STDP): Facilitates long-term weight modifications inspired by biological learning processes.
  • Tsodyks-Markram Short-Term Plasticity (STP): Accurately models both short-term depression and facilitation events, crucial for simulating real neuronal interactions.
• Configurable Transmission Delays: Features discrete synaptic delays, managed through efficient ring buffers, ensuring precise timing in network communications.
• Homeostatic Regulation: Implements synaptic scaling to stabilize activity within target frequency ranges (~10–12 Hz), promoting network robustness.
• Advanced Signal Encoding: Incorporates Time-To-First-Spike (TTFS) latency encoding, converting continuous inputs into accurately timed spike sequences that reflect real-world stimuli.
• Concurrency Support: Executes parallel processing by dividing contiguous slices efficiently (split_at_mut), effectively eliminating the need for mutex locks, thereby reducing thread contention during updates.

Performance Insights:

Extensive testing on configurations of 100 neurons across 4 CPU threads yielded:

• Execution Time: Approximately 2.63ms per simulation cycle.
• Network Dynamics: Achieved an Asynchronous Irregular (AI) firing state with average firing frequencies of ~12.6 Hz for excitatory neurons and ~10.0 Hz for inhibitory neurons, aligning with Dale's Principle.

This project serves as a valuable resource for researchers and developers interested in simulating complex neural networks efficiently, harnessing the power of modern multi-core processors.