Overview

Neural Substrate is a biophysical spiking neural network connected to a custom embodied environment. The model is grounded in primary literature on cortical and hippocampal microcircuitry rather than on artificial-network conventions: explicit cell types, membrane dynamics, Ca²⁺-driven plasticity, microglial pruning, and sleep-state replay are all modelled directly rather than approximated by scalar weights and gradient descent.

The work has two halves. The tissue model is a Rust + CUDA simulation of neurons, synapses, and structural plasticity. The environment is a 3D navigation world that the neural agent inhabits in place of a fixed dataset — sensory input arrives as ray-cast geometry and luminance fields, and actions are continuous heading deltas.

See the Neural Substrate research perspective for the framing question and the broader research stance behind this work.

Tissue model

  • Neurons

    Threshold-reset spike mechanism with explicit Ca²⁺ plateau, after-hyperpolarisation, HCN resonance, and slow adaptation. Two-compartment cells for burst-capable populations. Cell types include DG granule cells, CA3 and CA1 pyramidal cells, PV+ basket interneurons, and basal-ganglia populations (GPi, GPe, STN).

  • Synapses

    No scalar weight. Each synapse is a physical structure described by head volume, neck resistance, active zones, PSD-95 slots, and receptor occupancy. Effective transmission is the emergent product of those variables.

  • Plasticity and memory

    PKMζ/KIBRA commitment, NSF/PICK1 receptor trafficking, C1q/C3b microglial pruning with CD47-SIRPα veto and TREM2 regional suppression, perineuronal-net encapsulation, and per-synapse Wozniak stability with retrievability decay.

  • Sleep and replay

    Adenosine accumulation drives a VLPO-mediated sleep transition. Sleep-state sharp-wave-ripple replay and TREM2-modulated pruning run while the network is offline.

  • Spatial encoding

    Entorhinal grid cells across four hexagonal modules, plus place-field-tuned hippocampal cells, provide a spatial substrate that the embodied agent uses for navigation.

  • Compute

    CUDA kernels handle voltage updates, synaptic transmission, and structural updates. The simulation targets a Grace + Blackwell node: the GPU runs the fast voltage loop, the CPU runs structural plasticity per episode.

Environment

A 3D world with walls, food, water, and configurable scenarios. Two clients share the same world: the neural agent connects over a Unix socket using a RESET / ACTION / STATE protocol, and a human observer can attach via a WebSocket viewer. Sensory input is a 43-ray cast plus a 32×32 luminance field. Actions are continuous heading deltas with an idle option.

Cadence

Neural Substrate is research, not a product. Public updates here are posted roughly weekly with a short note on what changed and what was learned. Longer write-ups live under Papers and the Research note.

Project log

  • Provenance over behavior

    Published a methodology note on validating neural systems against relays. Behavior alone cannot establish that a system is doing the computation it appears to be doing — provenance, tracing every input back to its leaves, is the actual test. Generalizes beyond neural systems. Read the paper.

  • Project page expanded

    Neural Substrate's page now describes the tissue model and embodied environment directly. Weekly updates begin from this point.

  • A different kind of learning system

    Added a research note on stateful learning, internal drives, and adaptive behavior in the current Neural Substrate work.

  • Project record established

    Neural Substrate now has a project page, log source, and paper area for longer research documents.

Documents