Experiment 001 · Continuous-time learning
A learning system, in motion.
Forty points. Three internal states. Watch a network find its own boundary.
Learning live
01Decision boundary
Class AClass B
INPUT Y
−1.0INPUT X+1.0
TRAINING LOSS
0.2500
Evaluating initial network
ACCURACY
50%
20 / 40 points classified
02Learning curve
MEAN SQUARED ERROR
INITIALIZATION0.0 s
03Inside a neuron
HIDDEN 01
— v membrane— s adaptation— r refractory
0 perturbations · 0 accepted
Under the surface
More than an activation.
Each neuron has a membrane, a memory, and a moment to recover.
INDEPENDENT RECONSTRUCTIONThis is an independent reconstruction from Neuraxon’s description, not a port of Neuraxon’s source. The three coupled states follow the equations shown, integrated with fixed-step Euler updates. A small evolutionary search perturbs the 25 weights and keeps only changes that reduce loss on these 40 points. No backpropagation, libraries, or remote computation. The trace follows hidden neuron 01 as it cycles through the samples.
dv/dt = −v/τᵥ + input − αs
ds/dt = (v − s)/τₛ
dr/dt = −r/τᵣ + k·spike
output = sigmoid(v) · (1 − r)
τᵥ=.8 τₛ=2 τᵣ=1.2
α=.35 k=.12 Δt=.12
ds/dt = (v − s)/τₛ
dr/dt = −r/τᵣ + k·spike
output = sigmoid(v) · (1 − r)
τᵥ=.8 τₛ=2 τᵣ=1.2
α=.35 k=.12 Δt=.12