GPU-accelerated sparse liquid state machine for neuromorphic computing
Production-grade CUDA-accelerated sparse Liquid State Machine (LSM) with OU-SDE membrane dynamics, multi-lobe ensemble architecture, cuSPARSE mat-vec, and STDP covariance learning.
SparseBrain— configurable reservoir: N neurons, connectivity probability, Float16 sparse weights on GPU- Configurable input/output dimensions (
n_in,n_out) - OU-SDE dynamics:
dV = ((V_rest - V)/τ + I_rec + I_ext)dt + σ dW - cuSPARSE Float16 sparse mat-vec on GPU (fits 65k neurons in 16 GB VRAM)
- STDP covariance learning with eligibility traces
- 1000-tick rolling spike history buffer (circular, on-GPU)
EnsembleBrain— multi-lobe: multiple reservoirs with different time constants- Generic inhibition interface — caller provides a stress signal
using Pkg
Pkg.add("LiquidCortex")using LiquidCortex
# Create a 65,536-neuron sparse LSM lobe
brain = SparseBrain(20.0f0) # τ_m = 20ms, default n_in=14, n_out=16
# Or with custom dimensions
brain = SparseBrain(20.0f0; n_in=8, n_out=4)
# Or create the full 4-lobe ensemble (262,144 neurons)
ensemble = EnsembleBrain()
# Step the reservoir with an input vector
u = CUDA.zeros(Float32, 14)
step!(brain, u; inhibition=0.3f0)
# Read the output
output = get_output(brain)| Type / Function | Description |
|---|---|
SparseBrain(tau_m; n_in, n_out) |
Create a 65,536-neuron sparse reservoir lobe |
EnsembleBrain(; n_in, n_out) |
Create 4-lobe ensemble (262,144 neurons) |
step!(brain, u; inhibition, reflex_eta, ...) |
Execute one simulation timestep (see experimental kwargs) |
ensemble_step!(eb, u; inhibition, reflex_eta, reflex_signal, ...) |
Step all lobes and aggregate |
get_output(brain) |
Copy readout from GPU to CPU |
get_ensemble_output(eb) |
Copy aggregated readout |
compute_reservoir_covariance!(brain) |
Compute subsampled covariance matrix |
diagnostics(brain) |
Return diagnostic string |
ensemble_diagnostics(eb) |
Per-lobe diagnostic summary |
LiquidCortex is an experimental Julia package. Defaults are intentional:
| Keyword | Default / values | Meaning |
|---|---|---|
plasticity |
default :readout_only |
Frozen recurrent W; Hebbian W_out every 10 ticks |
opt-in :recurrent_stdp |
Experimental pair STDP on sparse edges every tick | |
opt-in :none |
No weight updates | |
recurrent_eta |
default 1f-4 |
Learning rate for :recurrent_stdp |
sync |
default true |
CUDA.synchronize() at end of step; host spike diagnostics only when true |
record_history |
default true |
Write spike history; if false, covariance helpers may see stale/incomplete history |
use_device_noise |
default false |
Host Gaussian noise upload; device RNG with host fallback if unavailable |
Recurrent reservoir weights are not trained under the default path. Requires CUDA.jl 6.x. Local verification and CI workflows use Julia 1.12.
dV = ((V_rest - V)/τ + W_rec·s(t) + W_in·x(t)) dt + σ dW
Discretized as Euler-Maruyama. Spike when V ≥ θ; reset to V_rest.
Ornstein & Uhlenbeck (1930); Maass, Natschläger & Markram (2002)
ΔW_ij = η (⟨s_i s_j⟩ - ⟨s_i⟩⟨s_j⟩)
Computed on a subsampled 8192-neuron window to avoid O(N²) blow-up.
Bi & Poo (1998); Hebb (1949)
Extracted from Eagle-Lander, a private neuromorphic GPU supervisor. The LSM core has been fully decoupled from domain-specific logic so it works with any time-series application.
Licensed under either of:
at your option.