Homeostasis¶

jaxfne includes a homeostatic excitability controller: a minimal computational control method that keeps each neuron's firing rate near a set-point by adapting an intrinsic excitability bias. It is one extra parameter per emitter, and switching it on does three useful things at once:

Eliminates hyperactivity — runaway/saturated units are pulled back down.
Eliminates hypoactivity — silent units are nudged up into a working range.
Models short-term adaptation — the slow rate trace gives spike-frequency adaptation for free.

It is a control method, not a claimed biological plasticity mechanism — a restoring controller on excitability, useful as a stabilizer and as an adaptation proxy. jaxfne is the mathematical backend; the controller's role in a model is whatever the configuration makes it.

The control law¶

Per neuron, a slow activity trace r tracks recent firing; a restoring bias g is added to the input current:

r ← decay · r + (1 − decay) · activity          # slow rate trace
g  = clip(k_gain · (r_star − r), g_min, g_max)   # over-active → g < 0, under-active → g > 0
I  ← I + g                                        # excitability bias into the input current

k_gain is the single dial. k_gain = 0 disables the controller — a clean null control for ablation experiments. The default k_gain = 1.0 is a gentle in-band nudge.

Built-in emitter (per-step kernel)¶

Enable it on the runtime; the built-in Izhikevich kernel applies the bias every step. After simulating, read the controller diagnostics off the model.

import jaxfne as jtfne
jtfne.enable_x64()

cfg = (
    jtfne.build_laminar_column(n=1000)
      .set_emitter("izhikevich", "cortical_eig")
      .probes(["spikes", "V_m"])
      .field(domain="laminar_column", conductivity="proxy")
      .runtime(
          duration_ms=1000.0, dt_ms=0.5, seed=0,
          enable_homeostasis=True,
          homeostasis_params={"k_gain": 1.0, "r_star": 0.01},
      )
)
model   = jtfne.construct(cfg)
signals = jtfne.simulate(model)

diag = model.last_homeostasis_diagnostics()   # {"g_bias": ..., "r_trace": ...}

homeostasis_params keys:

Key	Meaning
`r_star`	Target activity set-point the controller restores toward
`k_gain`	Restoring gain (`0` = disabled / null; default `1.0`)
`tau_r_ms`	Time constant of the slow rate trace
`alpha`	Trace update mixing factor
`g_min`, `g_max`	Clamp on the excitability bias
`r_max`	Activity-trace ceiling

model.last_homeostasis_diagnostics() returns the per-neuron g_bias and r_trace from the most recent enable_homeostasis=True run (or None if the last run had it off).

Jaxley emitter (outer-loop windowed)¶

The same control law wraps a Jaxley emitter, so you keep Jaxley's channel/morphology toolset and gain the homeostasis. Because Jaxley's integrate is a closed scan, the feedback runs at window cadence rather than per step, and the controller is parameterized by firing rate in Hz (which is dt-independent):

net = ...  # a Jaxley HH Network with recordings on the controlled compartments
bridge = jtfne.JaxleyBridge(model=net)

signals, diag = bridge.simulate_homeostatic(
    duration_ms=400.0,
    base_current_nA=0.01,
    target_rate_hz=40.0,    # set-point in Hz
    k_gain=0.2,             # 0 = null
    window_ms=20.0,
    return_diagnostics=True,
)
# diag["g_bias"], diag["r_trace"], diag["rate_hz"] -> [n_windows, n_cells]

The injected current is hard-bounded (current_clip_nA) and finiteness is verified (strict_finite=True raises rather than masking) — the controller keeps the implicit solver in a finite, stable regime even under heterogeneous drive. See JaxleyBridge.simulate_homeostatic in the Bridges API.

Stay in the monotonic f-I band (Jaxley path)

Jaxley single-compartment HH has a non-monotonic f-I curve (high current → depolarization block). Keep base_current_nA and the bias range inside the monotonic band (roughly 0.002–0.02 nA for the default compartment), so the controller operates where firing rate increases with current.

Stability¶

Both paths are float32-safe: the built-in kernel hard-bounds its state (v/u/synaptic variables) so the dynamics stay finite even under extreme drive, and the Jaxley path hard-bounds the injected current and checks finiteness. The controller therefore doubles as a numerical stabilizer.

Using it as evidence¶

Because k_gain = 0 is a true null, the controller is built for clean comparisons: run with k_gain = 0 and k_gain > 0, hold everything else fixed, and attribute the difference (rate spread collapsing toward the set-point, silent units recovering) to the controller. Report the null alongside the result.