MH-FLOCKE MH-FLOCKE
HomeDocsGitHubBlogPaperYouTubeReddit𝕏

Purkinje Compartments

Purkinje Multi-Compartment Layer

Purkinje cells in the biological cerebellum have complex dendritic trees with distinct computational compartments. MH-FLOCKE models this with a 3-compartment neuron: soma, apical dendrite, and basal dendrite.

Three Compartments

  • Soma (τ = 15ms) — Output compartment. Fires simple spikes for DCN inhibition.
  • Apical Dendrite (τ = 80ms) — Slow integration of parallel fiber (GrC) input. The main plastic site for LTD/LTP learning.
  • Basal Dendrite (τ = 5ms) — Fast response to climbing fiber input. Triggers calcium spikes that gate LTD in the apical compartment.

Calcium Dynamics

When a climbing fiber fires (basal input), calcium floods the apical compartment. Calcium persists for ~12 steps (vs 1 step for the CF itself), creating a wider learning window. This is biologically more accurate than binary CF gating.

calcium += cf_input × 0.8
calcium *= 0.85  (decay per step)
LTD gate = calcium (used for PF→PkC weight update)

Complex Spikes

When calcium exceeds threshold (1.5), a complex spike is generated — a burst of output that strongly inhibits the DCN. This is the Purkinje cell’s strongest signal.

References

  • Rancz et al. (2007). Dendritic spikes in Purkinje cells. Neuron
  • D’Angelo (2025). Linking cellular phenomena to brain architecture. Trends in Neurosciences

API Reference

PurkinjeCompartmentLayer(config: PurkinjeConfig)

step(pf_input, cf_input)

Update all compartments: apical integrates PF, basal responds to CF, soma outputs activity. Calcium dynamics for LTD gating.

get_ltd_gate() → Tensor[n_pkc]

Current calcium levels for LTD/LTP decision (0 = LTP window, >0.05 = LTD active).

get_compartment_state() → dict

Returns soma, apical, basal voltages, calcium, complex_spike flags.

PurkinjeConfig

n_neurons: auto (2 × n_actuators)
tau_soma: 15ms   tau_apical: 80ms   tau_basal: 5ms