DavidVivancos/Neuraxon

What it solves

Neuraxon 2.0 is a bio-inspired computational framework designed to overcome the limitations of traditional artificial neural networks (ANNs) and spiking neural networks (SNNs), such as discrete time steps, static topologies, and catastrophic forgetting. It provides a blueprint for neural growth and computation that allows for continuous real-time learning and biological plausibility.

How it works

Unlike the classic perceptron, Neuraxon uses trinary logic (+1 excitatory, 0 neutral, -1 inhibitory) and operates in continuous time. It implements a complex system of biological mechanisms including:

• Multi-timescale Plasticity: Synapses have three dynamic weights (fast, slow, and meta) and can form, collapse, or reconnect.
• Neuromodulation: Four neuromodulators (DA, 5-HT, ACh, NA) with nine receptor subtypes regulate network activity.
• Unified Pipeline: A 4-step process involving Time Warping (ChronoPlasticity), Dynamic Decay (DSN), CTSN Complemented State, and Astrocyte-Gated Multi-timescale Plasticity (AGMP).
• Structural Plasticity: Hidden neurons can die and connections can be rewired using Watts-Strogatz small-world topology.
• Hybridization: Integration with Aigarth Intelligent Tissue for evolutionary mutation and selection.

Who it’s for

This project is intended for researchers in neuroscience, artificial life, and AI developers focusing on embodied robotic control, real-time temporal pattern recognition, and the development of energy-efficient neuromorphic hardware.

Highlights

• Continuous-time learning: No separate training and inference phases; the network adapts instantly to streaming data.
• Biological plausibility: Includes dendrites, silent synapses, and homeostatic regulatory loops (MSTH).
• Trinary state logic: Uses a more nuanced state system than binary spiking.
• Evolvability: Supports population-level evolution of network structure and parameters via Aigarth hybrid.
• Cuda support: Includes CuNxon Cuda kernels for accelerated computation.