Skip to content

Home

Jump to bottom
Raul Montoya Cardenas edited this page Jul 29, 2026 · 3 revisions

LiquidCortex.jl

Wiki hero

Generated with Grok Build: Grok 4.5 · xAI Imagine (/imagine)

GPU-accelerated sparse Liquid State Machine for neuromorphic computing.

Version: 0.2.0 | Language: Julia 1.10–1.12 | License: MIT OR Apache-2.0

LiquidCortex provides a production-grade CUDA LSM with OU-SDE membrane dynamics, multi-lobe ensemble architecture, cuSPARSE Float16 mat-vec, and STDP covariance learning. The core is domain-agnostic: callers supply generic input vectors and an optional inhibition signal.

What This Package Covers

  • SparseBrain — 65,536-neuron sparse CUDA reservoir lobe (1% connectivity, Float16 weights)
  • EnsembleBrain — 4 lobes × 65,536 = 262,144 neurons with τ_m ∈ {10, 25, 50, 100} ms
  • OU-SDE / LIF dynamics with refractory masking and adaptive threshold inhibition
  • STDP eligibility traces + Hebbian W_out updates; subsampled reservoir covariance
  • Reference LSM — 2,048-neuron dense CUDA reservoir for rapid prototyping
  • Clean CPU load: types and API defined without a GPU; allocations deferred until CUDA is available
  • Optional Sentry runtime exception capture (SENTRY_DSN)

Start Here

Page Description
Getting Started Install, hardware needs, first SparseBrain / EnsembleBrain steps
Project Structure Repo layout, module entry, _cuda_available
Core Architecture LSM paradigm, SparseBrain vs Reference LSM
SparseBrain and EnsembleBrain 65k-lobe design, ensemble weights, VRAM budget
Inhibition and Reflex Gating Global inhibition, flash-learning on Fast lobe
STDP and Covariance Learning rule, eligibility traces, covariance subsample
Reference LSM 2,048-neuron dense prototype API
Public API Exported symbols overview
Neural Execution API Constructors, step!, outputs, diagnostics
Input Interface Generic CuVector{Float32} contract
Infrastructure CI, Codecov, Sentry automation
Testing and Examples Test suite and standalone example
Glossary Terms, constants, and code pointers

Ecosystem Role

LiquidCortex is the cortical reservoir layer in the Limen-Neural stack: it turns continuous time-series into high-dimensional spike-driven readouts that downstream systems can gate, size, or train against.

Layer Responsibility Entity
Encoding Continuous → spikes axon-encoder (Rust)
Reservoir / Cortex Sparse LSM, multi-timescale dynamics LiquidCortex.jl
Strategy / Control Confidence gate, Kelly sizing, backtest DendriteTrader.jl
Execution / Muscle HFT loops, IPC, venues corpus-ipc (Rust)

Public API (at a glance)

Type / Function Description
SparseBrain(tau_m; n_in, n_out, name) 65,536-neuron sparse lobe
EnsembleBrain(; n_in, n_out) 4-lobe ensemble (262,144 neurons)
step!(brain, u; inhibition, reflex_eta) One SparseBrain timestep
ensemble_step!(eb, u; inhibition, reflex_eta, reflex_signal) Step all lobes + aggregate
get_output / get_ensemble_output GPU → CPU readout copy
compute_reservoir_covariance! Subsampled 8192-neuron covariance
diagnostics / ensemble_diagnostics Status strings

Provenance

Extracted from Eagle-Lander, a private neuromorphic GPU supervisor. Domain-specific market/mining telemetry was removed in PR #12 so the LSM works with any time-series application.

License

Dual-licensed under MIT or Apache-2.0 at your option.

Last updated: July 28, 2026 Updated by: Grok Build: Grok 4.5 Package tip reference: 4e2698c (main, through PR #34)

Clone this wiki locally