Changes here apply from the next trial. Reset preset restores the preset's own settings.
Off is a deliberately unbalanced demonstration: local LTP continues at its usual rate but nothing balances it outside error-driven CF events, so predictions drift upward until overprediction drives enough CF events to cancel LTP.
The local LTP rate constant is calibrated to this baseline (LTP = η · p₀ · eligibility), so any of these settings is balanced on average when the error is zero.
With late feedback, error-modulated CF events arrive when no synapse is eligible, so they cause no LTD and no learning. You can also press Trigger CF during and after the eligibility glow.
The default seed is fixed, so a lecture can be repeated exactly. Reset preset returns to the default seed. Default seeds: A 322, B 22, C 31.
N, → or Page Down advance one event. P play or pause. R repeat trial. M new movement. 1 2 3 select a PC. C trigger CF at the selected PC. Shift+R reset preset. I this panel. Esc close.
GC features are fixed products of MF inputs, for example GC6 = elbow position × biceps tension. PC prediction for cell i is Σj wij·GCj, using the weights at the moment of prediction. The prediction is saved and never recomputed.
Eligibility of synapse GCj→PCi: e = GCj·exp(−t/300 ms), treated as expired once the factor falls below 0.15 (≈570 ms).
Plasticity opportunity every 100 ms of model time: local LTP adds η·p₀·e at every eligible synapse; each PC independently draws a CF event with probability p. A CF event at PCi subtracts η·e from that PC's eligible synapses only. Weights are kept in [0, 1].
CF probability is the baseline p₀ = 0.20, except in the feedback-modulated window (400 ms after feedback arrives), where pi = clip(p₀ − k·ri, 0, 1) with k = 2.5 and ri = actual − saved prediction. Overprediction (r < 0) means more CF events; underprediction (r > 0) means fewer, so LTP wins. When r = 0, LTP and CF-driven LTD balance on average. η = 0.05.