See the Neural Substrate project page for the technical breakdown, weekly project log, and document links.
The question
Can a biologically grounded neural system accumulate useful state through continuous interaction with an environment? The target is not just task improvement, but preservation of dynamics, structure, and operating continuity over time.
Current architecture
The current system combines adaptive neuron dynamics, physical synapse state, structural plasticity, CUDA kernels, and a navigation-scale agent connected to a custom 3D simulation through a Unix socket.
Rust/CUDA tissue model with adaptive neurons and physical synapse variables.
Hippocampal, cortical, basal-ganglia, thalamic, cerebellar, olfactory, and neuromodulatory populations.
A custom navigation simulator provides the current body and world interface.
Agent control over Unix socket, browser visualization, ray depth, bifocal luminance, and scent inputs.
Implemented mechanisms
Current code includes calcium plateaus, physical synapses, PKM zeta and KIBRA commitment logic, C1q/C3b pruning signals, CD47 veto behavior, adenosine and sleep dynamics, replay episodes, perineuronal-net encapsulation, and entorhinal grid-cell encoding.
Predictions
Two testable predictions about ring-attractor dynamics in head-direction systems. Both are framed against the fly ellipsoid body (EB), where the underlying circuit is public, but the predictions apply to any spiking ring-attractor model.
Symmetric two-bump separation of mechanism. The standard single-bump suppression assay — stimulate a second site while a reference bump exists — cannot distinguish a structural attractor from history-dependent or short-term-plasticity-based suppression. Both mechanisms pass the asymmetric protocol. A symmetric two-bump protocol with equal drive to two sites and no temporal precedence advantage separates them. Prediction: history-independent winner-take-all dynamics in a ring attractor require a non-depressing excitatory-to-global-inhibitor synapse. If a candidate circuit exhibits short-term depression on this pathway, it should fail the symmetric protocol while passing the asymmetric one.
Granularity of the inhibitory population. Single-bump stability requires the global-inhibitory population to be fine-grained enough relative to the ring. If the inhibitory coverage is too coarse, the bump fragments into multi-peak configurations in the angular gaps between inhibitory tiles. Framed here as an observation rather than a discovery; novelty relative to attractor theory is uncertain.
Methodology
A separate methodology paper describes the validation approach used in this work: behavior alone cannot distinguish a genuine computation from a relay, so provenance — tracing every input to its leaves — becomes the actual test. The paper generalizes beyond neural systems.
Current limits
This is active research, not a product. Some pathways are complete enough to test, while others remain scaffolds, partial biological approximations, or integration work between the substrate and the simulator.