Dynamic Input Conductances (DICs)

• Dynamic Input Conductances (DICs) are voltage-dependent, time-scale-specific effective conductances that consolidate ion channel effects into fast, slow, and ultra-slow feedback components.
• They are applied in optogenetics and control-theoretic models, where fluctuating conductances simulate synaptic noise and stabilize learned dynamical systems.
• DIC frameworks bridge biophysical neuron models with computational reconstruction techniques, enabling high reproducibility and accurate spike-based inference.

to=arxiv_search 大发时时彩怎么.json code {"query":"Dynamic Input Conductances ICNDM CoDyPs spike times conductance-based neuron models", "max_results": 10} Dynamic Input Conductances (DICs) are time-varying conductance descriptions used to characterize how inputs are transmitted into state changes in dynamical systems. In conductance-based neuroscience, DICs are voltage-dependent, timescale-specific effective conductances that aggregate the influence of ion channels into fast, slow, and ultra-slow feedback components; they are also used experimentally as imposed fluctuating conductances that emulate synaptic bombardment. In more recent machine-learning and control settings, the same idea is used conceptually to describe how the responsiveness or gain from inputs to states evolves over time while remaining stable or well-behaved (Brandoit et al., 16 Sep 2025, Neef et al., 2013, Pouladi, 7 Jul 2026).

1. Definitions and conceptual scope

In the conductance-based setting, DICs are defined as “voltage-dependent, timescale-specific effective conductances” that reduce the influence of many ion channels to three interpretable components: the fast dynamic input conductance $g_f(V)$, the slow dynamic input conductance $g_s(V)$, and the ultra-slow dynamic input conductance $g_u(V)$ (Brandoit et al., 16 Sep 2025). These quantities describe how the total membrane current responds to small perturbations in voltage on distinct timescales associated with spike initiation and upstroke, spike repolarization and interspike-interval dynamics, and adaptation, burst envelope, and interburst dynamics.

This use of “input conductance” is not limited to static membrane slopes. The designation “dynamic” reflects the fact that gating kinetics, through $\tau_X(V)$ and $X_\infty(V)$, are incorporated directly rather than being replaced by instantaneous activation curves alone (Brandoit et al., 16 Sep 2025). In this sense, DICs summarize the feedback architecture of excitability rather than merely the passive input resistance of a cell.

A second, experimentally oriented meaning appears in optogenetic stimulation. Continuous dynamic photostimulation (CoDyPs) is described as an optogenetic implementation of DICs: fluctuating blue light $L(t)$ drives a channelrhodopsin-mediated conductance $g_{\text{ChR}}(t)$ that acts as a designed, time-varying membrane conductance analogous to imposed synaptic noise (Neef et al., 2013). Here DICs are not only descriptors of intrinsic dynamics, but imposed conductance waveforms with specified statistics.

A third usage is explicitly conceptual. The Input-Contraction Neural Differential Model (ICNDM) treats DICs as a way to think about how a controlled dynamical system’s “responsiveness” or “gain” from inputs to states evolves over time while remaining stable or well-behaved (Pouladi, 7 Jul 2026). In that formulation, the term does not denote classical neuronal membrane conductances; it denotes an input-aware geometry and bounded input-to-state gain in a learned continuous-time model.

2. Conductance-based formulation and timescale decomposition

The starting point for the formal DIC framework is a Hodgkin–Huxley-type conductance-based model,

$$C \frac{dV}{dt} + g_\text{leak}(V - E_\text{leak}) = - \sum_{i \in \mathcal{I}} \bar{g}_{i} m_{i}^{p_{i}}(V,t)\, h_{i}^{q_{i}}(V,t)\,(V - E_{i}) + I_{\text{ext}}(t),$$

with ionic currents

$$I_i(V,t) = \bar{g}_i\, m_i^{p_i}(V,t)\, h_i^{q_i}(V,t)\, (V - E_i),$$

and first-order gating kinetics

$$\tau_X(V)\,\frac{dX}{dt} = X_\infty(V) - X, \quad X\in\{m_i,h_i\}.$$

Within this framework, the DICs are defined by a timescale decomposition of the membrane feedback terms (Brandoit et al., 16 Sep 2025):

$g_s(V)$0

The weighting functions $g_s(V)$1 and $g_s(V)$2 assign each gating variable’s influence across fast, slow, and ultra-slow timescales according to its voltage-dependent time constant. The DICs are normalized by $g_s(V)$3, so they are dimensionless ratios relative to leak. The sign convention is also explicit: positive DICs correspond to regenerative, destabilizing contributions, whereas negative DICs correspond to stabilizing contributions (Brandoit et al., 16 Sep 2025).

For a fixed conductance-based model, the three DIC curves admit a sensitivity-matrix representation,

$g_s(V)$4

where $g_s(V)$5 is the vector of maximal conductances plus leak. This expresses the DICs as aggregated feedback gains obtained from the entire channel repertoire.

A further reduction is obtained by evaluating the DICs at threshold voltage. The total conductance curve is defined as

$g_s(V)$6

and the threshold voltage $g_s(V)$7 is taken as the first decreasing zero,

$g_s(V)$8

The three scalars $g_s(V)$9, $g_u(V)$0, and $g_u(V)$1 then summarize the spontaneous firing regime in a three-dimensional space (Brandoit et al., 16 Sep 2025).

Component	Timescale	Principal role
$g_u(V)$2	Fast	Spike initiation and upstroke
$g_u(V)$3	Slow	Pattern selection, including spiking vs bursting
$g_u(V)$4	Ultra-slow	Burst envelope, interburst dynamics, adaptation

The slow component is identified as the primary indicator of firing regime: $g_u(V)$5 corresponds to tonic spiking, whereas $g_u(V)$6 promotes bursting, with the transition occurring near $g_u(V)$7 (Brandoit et al., 16 Sep 2025). The ultra-slow component modulates burst structure and adaptation, and weak ultra-slow negative feedback can produce a “fast spiking” regime in the dopaminergic neuron model discussed in that work.

3. Experimental realization through continuous dynamic photostimulation

CoDyPs implements DICs optically rather than electrically. Neurons express a light-gated cation channel, primarily ChIEF, and are illuminated with fluctuating blue light $g_u(V)$8. The channelrhodopsin converts the light stimulus into a time-varying conductance $g_u(V)$9, producing a photocurrent

$\tau_X(V)$0

with reversal potential near $\tau_X(V)$1 mV (Neef et al., 2013). The stated aim is to mimic in-vivo-like input fluctuations noninvasively.

The stimulus is typically designed as an Ornstein–Uhlenbeck process with specified mean, variance, and correlation time. Over the operating range used in the study, ChIEF behaves as a linear time-invariant element whose impulse response function $\tau_X(V)$2 maps light to conductance or current:

$\tau_X(V)$3

The impulse response is estimated by frequency-domain division,

$\tau_X(V)$4

This provides a calibrated light–conductance transfer function (Neef et al., 2013).

The measured transfer characteristics are central to the DIC interpretation. The IRF exhibits a brief delay of approximately $\tau_X(V)$5–$\tau_X(V)$6 $\tau_X(V)$7s, followed by an effective single-exponential decay with $\tau_X(V)$8–$\tau_X(V)$9 ms. Chirp experiments show that the current behaves as a first-order low-pass filter with cutoff approximately $X_\infty(V)$0 Hz,

$X_\infty(V)$1

As a consequence, frequencies up to roughly $X_\infty(V)$2 Hz are transferred with reasonable fidelity, whereas higher frequencies are attenuated (Neef et al., 2013).

The experimental paper reports strong reproducibility. In voltage-clamp experiments in HEK cells, CoDyPs-induced current waveforms can be predicted from the measured light sequence by convolution with the IRF, with correlation between predicted and measured current of approximately $X_\infty(V)$3–$X_\infty(V)$4 across conditions. Trial-to-trial current correlations are typically $X_\infty(V)$5–$X_\infty(V)$6 for $X_\infty(V)$7 ms, and trial-to-trial deviations are below $X_\infty(V)$8 pA (Neef et al., 2013).

At the level of spiking, CoDyPs drives neurons into a fluctuation-driven regime. Spike trains appear irregular, but repeated presentation of the same light sequence yields highly reproducible spike patterns across trials and even across hours. The paper describes up to $X_\infty(V)$9 hours of repeated CoDyPs stimulation with stable spike patterns (Neef et al., 2013). This irregular-yet-reliable regime is precisely the one for which imposed DICs are used to study dynamic response properties.

4. DICs as an intermediate representation for spike-based reconstruction

A recent extension uses DICs as a bridge from observed spike times to populations of conductance-based models. In that framework, DICs at threshold condense high-dimensional conductance variability into three scalar constraints that preserve firing regime and major activity statistics (Brandoit et al., 16 Sep 2025). The central pipeline is

$L(t)$0

where $L(t)$1 is the spike-time sequence and $L(t)$2 is a maximal-conductance vector compatible with the observed activity.

The spike-time encoder operates on interspike intervals and second differences, standardizes these features, augments them with positional encodings, and processes them with Transformer-style multi-head self-attention blocks and a self-attention pooler. The decoder contains four heads: a main DIC regression head predicting $L(t)$3, a classification head for spiking versus bursting, an activity-metrics head, and an uncertainty head for heteroscedastic loss weighting (Brandoit et al., 16 Sep 2025). Fast DIC is treated separately: for dataset generation it is fixed at a sufficiently negative value ensuring excitability, while inference focuses on the slow and ultra-slow components.

Reconstruction of maximal conductances proceeds by compensation against the sensitivity matrix. Because DIC constraints are far fewer than conductance parameters, the inverse problem is underdetermined and naturally supports degeneracy. The paper therefore generates “twin” neuron models by sampling part of the conductance vector from a broad prior and solving linear systems that enforce the target DIC constraints at threshold. For models in which the sensitivity matrix depends nonlinearly on conductances, an iterative compensation scheme is used, updating the sensitivity matrix and re-solving until the residual

$L(t)$4

becomes sufficiently small (Brandoit et al., 16 Sep 2025). The paper states that typically five iterations reduce the residual by approximately $L(t)$5 while preserving degeneracy.

The reported performance on the stomatogastric ganglion model is explicit: firing-class accuracy is $L(t)$6, spiking-frequency mean absolute error is $L(t)$7 Hz versus dataset standard deviation $L(t)$8 Hz, and DIC prediction MAEs are $L(t)$9 for $g_{\text{ChR}}(t)$0 and $g_{\text{ChR}}(t)$1 for $g_{\text{ChR}}(t)$2 overall (Brandoit et al., 16 Sep 2025). By regime, $g_{\text{ChR}}(t)$3 MAE is $g_{\text{ChR}}(t)$4 in spiking and $g_{\text{ChR}}(t)$5 in bursting, while $g_{\text{ChR}}(t)$6 MAE is $g_{\text{ChR}}(t)$7 in spiking and $g_{\text{ChR}}(t)$8 in bursting. The larger spiking errors are attributed to DIC degeneracy rather than failure to capture activity, because tonic spiking occupies a one-dimensional manifold in DIC space and many $g_{\text{ChR}}(t)$9 pairs yield nearly identical spiking patterns.

This degeneracy is not treated as a defect to be eliminated. Rather, DICs are used to structure a degenerate solution space in which maximal conductances can vary substantially while spike statistics remain similar (Brandoit et al., 16 Sep 2025).

5. Control-theoretic generalization in learned dynamical systems

In the ICNDM framework, DICs are used conceptually to interpret input responsiveness in controlled continuous-time models. The learned system has the form

$$C \frac{dV}{dt} + g_\text{leak}(V - E_\text{leak}) = - \sum_{i \in \mathcal{I}} \bar{g}_{i} m_{i}^{p_{i}}(V,t)\, h_{i}^{q_{i}}(V,t)\,(V - E_{i}) + I_{\text{ext}}(t),$$0

where $$C \frac{dV}{dt} + g_\text{leak}(V - E_\text{leak}) = - \sum_{i \in \mathcal{I}} \bar{g}_{i} m_{i}^{p_{i}}(V,t)\, h_{i}^{q_{i}}(V,t)\,(V - E_{i}) + I_{\text{ext}}(t),$$1 is the neural vector field and $$C \frac{dV}{dt} + g_\text{leak}(V - E_\text{leak}) = - \sum_{i \in \mathcal{I}} \bar{g}_{i} m_{i}^{p_{i}}(V,t)\, h_{i}^{q_{i}}(V,t)\,(V - E_{i}) + I_{\text{ext}}(t),$$2 is a neural input encoder that maps the raw control input $$C \frac{dV}{dt} + g_\text{leak}(V - E_\text{leak}) = - \sum_{i \in \mathcal{I}} \bar{g}_{i} m_{i}^{p_{i}}(V,t)\, h_{i}^{q_{i}}(V,t)\,(V - E_{i}) + I_{\text{ext}}(t),$$3 to a latent representation (Pouladi, 7 Jul 2026). The paper explicitly interprets the encoder as part of the mechanism determining the “dynamic conductance” from inputs to states, because it shapes the effective input channels before they affect $$C \frac{dV}{dt} + g_\text{leak}(V - E_\text{leak}) = - \sum_{i \in \mathcal{I}} \bar{g}_{i} m_{i}^{p_{i}}(V,t)\, h_{i}^{q_{i}}(V,t)\,(V - E_{i}) + I_{\text{ext}}(t),$$4.

Stability is enforced through a learned Riemannian metric,

$$C \frac{dV}{dt} + g_\text{leak}(V - E_\text{leak}) = - \sum_{i \in \mathcal{I}} \bar{g}_{i} m_{i}^{p_{i}}(V,t)\, h_{i}^{q_{i}}(V,t)\,(V - E_{i}) + I_{\text{ext}}(t),$$5

and the contraction matrix

$$C \frac{dV}{dt} + g_\text{leak}(V - E_\text{leak}) = - \sum_{i \in \mathcal{I}} \bar{g}_{i} m_{i}^{p_{i}}(V,t)\, h_{i}^{q_{i}}(V,t)\,(V - E_{i}) + I_{\text{ext}}(t),$$6

with

$$C \frac{dV}{dt} + g_\text{leak}(V - E_\text{leak}) = - \sum_{i \in \mathcal{I}} \bar{g}_{i} m_{i}^{p_{i}}(V,t)\, h_{i}^{q_{i}}(V,t)\,(V - E_{i}) + I_{\text{ext}}(t),$$7

The core condition is

$$C \frac{dV}{dt} + g_\text{leak}(V - E_\text{leak}) = - \sum_{i \in \mathcal{I}} \bar{g}_{i} m_{i}^{p_{i}}(V,t)\, h_{i}^{q_{i}}(V,t)\,(V - E_{i}) + I_{\text{ext}}(t),$$8

which yields incremental exponential convergence for trajectories driven by the same time-varying input (Pouladi, 7 Jul 2026).

The more directly DIC-like statement is the input-to-state contraction bound. If the vector field is Lipschitz in inputs,

$$C \frac{dV}{dt} + g_\text{leak}(V - E_\text{leak}) = - \sum_{i \in \mathcal{I}} \bar{g}_{i} m_{i}^{p_{i}}(V,t)\, h_{i}^{q_{i}}(V,t)\,(V - E_{i}) + I_{\text{ext}}(t),$$9

then trajectories driven by different inputs satisfy

$$I_i(V,t) = \bar{g}_i\, m_i^{p_i}(V,t)\, h_i^{q_i}(V,t)\, (V - E_i),$$0

The paper identifies the gain factor $$I_i(V,t) = \bar{g}_i\, m_i^{p_i}(V,t)\, h_i^{q_i}(V,t)\, (V - E_i),$$1 as a rigorous input-to-state sensitivity bound (Pouladi, 7 Jul 2026). This suggests a control-theoretic analogue of dynamic input conductance: input differences are conducted into state differences through a gain shaped by metric bounds, contraction rate, and input Lipschitz constant.

Training combines prediction, contraction regularization, and input-sensitivity regularization,

$$I_i(V,t) = \bar{g}_i\, m_i^{p_i}(V,t)\, h_i^{q_i}(V,t)\, (V - E_i),$$2

The contraction penalty is based on the largest eigenvalue of $$I_i(V,t) = \bar{g}_i\, m_i^{p_i}(V,t)\, h_i^{q_i}(V,t)\, (V - E_i),$$3, and the input-sensitivity term penalizes $$I_i(V,t) = \bar{g}_i\, m_i^{p_i}(V,t)\, h_i^{q_i}(V,t)\, (V - E_i),$$4 (Pouladi, 7 Jul 2026). In experimental evaluations, ICNDM achieves a Duffing-oscillator $$I_i(V,t) = \bar{g}_i\, m_i^{p_i}(V,t)\, h_i^{q_i}(V,t)\, (V - E_i),$$5-step rollout MSE of $$I_i(V,t) = \bar{g}_i\, m_i^{p_i}(V,t)\, h_i^{q_i}(V,t)\, (V - E_i),$$6, compared with $$I_i(V,t) = \bar{g}_i\, m_i^{p_i}(V,t)\, h_i^{q_i}(V,t)\, (V - E_i),$$7 for a standard NODE, and a Van der Pol $$I_i(V,t) = \bar{g}_i\, m_i^{p_i}(V,t)\, h_i^{q_i}(V,t)\, (V - E_i),$$8-step rollout MSE of $$I_i(V,t) = \bar{g}_i\, m_i^{p_i}(V,t)\, h_i^{q_i}(V,t)\, (V - E_i),$$9, compared with $$\tau_X(V)\,\frac{dX}{dt} = X_\infty(V) - X, \quad X\in\{m_i,h_i\}.$$0 for a standard NODE. The same study also reports close tracking on a Permanent Magnet Synchronous Motor drive system and graceful degradation under injected Gaussian input noise (Pouladi, 7 Jul 2026).

A common misconception is to treat these control-theoretic DICs as identical to the conductance-based neuronal formalism. The sources support a narrower statement: ICNDM is a principled, contraction-theoretic implementation of dynamic conductance as an analogy for bounded input-to-state responsiveness, not a reformulation of the classical neuronal DIC definitions (Pouladi, 7 Jul 2026).

6. Applications, limitations, and open problems

Across these literatures, DICs serve as a compact interface between complex internal dynamics and interpretable input responsiveness. In biophysical modeling, they reduce many channels and gating variables to three feedback gains that organize excitability, spiking, bursting, and adaptation (Brandoit et al., 16 Sep 2025). In optogenetics, they provide a way to impose fluctuating conductances with controlled mean, variance, autocorrelation, and power spectrum, enabling long-term and large-scale studies of fluctuation-driven spiking without patch electrodes (Neef et al., 2013). In learned control models, they motivate architectures that bound input-to-state gain while preserving long-horizon predictive accuracy (Pouladi, 7 Jul 2026).

The limitations are domain-specific. In the spike-based reconstruction framework, DICs are evaluated at threshold only; subthreshold and suprathreshold waveform details are not directly constrained, responses to external input are not modeled, and the model structure must be fixed in advance (Brandoit et al., 16 Sep 2025). In CoDyPs, the conductance is low-pass filtered by ChIEF kinetics with an $$\tau_X(V)\,\frac{dX}{dt} = X_\infty(V) - X, \quad X\in\{m_i,h_i\}.$$1 ms time constant and approximately $$\tau_X(V)\,\frac{dX}{dt} = X_\infty(V) - X, \quad X\in\{m_i,h_i\}.$$2 Hz cutoff, the channel is a mixed Na$$\tau_X(V)\,\frac{dX}{dt} = X_\infty(V) - X, \quad X\in\{m_i,h_i\}.$$3/K$$\tau_X(V)\,\frac{dX}{dt} = X_\infty(V) - X, \quad X\in\{m_i,h_i\}.$$4/H$$\tau_X(V)\,\frac{dX}{dt} = X_\infty(V) - X, \quad X\in\{m_i,h_i\}.$$5 cation channel rather than a physiological glutamatergic or GABAergic synapse, and the conductance is distributed wherever channelrhodopsin is expressed rather than being synapse-specific (Neef et al., 2013). In ICNDM, global contraction can be conservative, the learned metric is parameterized as a function of state rather than explicitly as $$\tau_X(V)\,\frac{dX}{dt} = X_\infty(V) - X, \quad X\in\{m_i,h_i\}.$$6, and guarantees hold over an operational domain $$\tau_X(V)\,\frac{dX}{dt} = X_\infty(V) - X, \quad X\in\{m_i,h_i\}.$$7 rather than universally (Pouladi, 7 Jul 2026).

Several future directions are stated directly in the sources. The spike-based framework points to multi-voltage DIC constraints, stimulus-response datasets with $$\tau_X(V)\,\frac{dX}{dt} = X_\infty(V) - X, \quad X\in\{m_i,h_i\}.$$8, synaptic DICs for network modeling, adaptive neuromodulation and closed-loop control, and broader transfer across cell types (Brandoit et al., 16 Sep 2025). The optogenetic literature suggests population studies with correlated light-driven conductances and in vivo naturalistic perturbations (Neef et al., 2013). The control literature suggests extending contraction-based learning toward richer region-specific or explicitly input-dependent geometries (Pouladi, 7 Jul 2026).

Taken together, these developments establish DICs as a unifying language for time-varying conductance and gain structure. In neurons, DICs provide a mechanistically grounded decomposition of excitability into fast, slow, and ultra-slow feedback. In experiments, they can be imposed optically as reproducible fluctuating conductances. In learned controlled dynamics, they motivate input-aware, stability-certified representations of how external signals are conducted into state evolution.