acceptodds
Under review as a conference paper at ICLR 2027

Retracing Hodgkin and Huxley: State Recovery Does Not Certify Mechanism

Abstract

Predicting observed dynamics does not establish recovery of the underlying physical mechanism. Can machine learning retrace the hidden-state reasoning behind the Hodgkin–Huxley (HH) model? We train structured latent models on simulated current and voltage, withholding gate identities and trajectories from training and model selection. We then test response prediction, state recovery, protocol transfer, and agreement with HH dynamics. Prediction error and its cross-seed spread both drop sharply at three latent dimensions under the tested protocols, while gate recovery under new protocols improves through five to six coordinates. State recovery depends on which observations the chart uses. Observed voltage improves current-clamp decoding relative to freely predicted voltage. Under voltage clamp, adding latent state to command voltage raises m-state R² from 0.976 to above 0.99, yet the transported field disagrees with HH on identical smooth samples. Known invertible HH coordinates achieve high fast-m field agreement under the same audit procedure. An exact HH identity decomposes the discrepancy into time-scale-weighted state error and a residual in the transported field; these terms can cancel or reinforce. These findings concern the tested models and charts. They support evaluating state and dynamics recovery separately, including chart inputs and transported-field agreement across interventions.

Primary area:
learning on time series and dynamical systems
Keywords:
Mechanistic Interpretability, Latent State Recovery, Dynamical Systems, Scientific Machine Learning, Hodgkin–Huxley Model, Neural Dynamics

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_czpROpenReview ↗
Rating
2
Confidence
2
Informativeness
1
Presentation
1
Soundness
2
Read review
Paper summary

The authors used Hodgkin-Huxley model as a test-bed to see if learned model that is able to predict observables can find the physically meaningful latent states. The motivation is clear and relevant, and the test case is a good one but the current presentation is hard to follow for me.

Critical strengths
  • The Hodgkin-Huxley model is one of the most successful story in modeling that find the mechanism. So it is a great idea to see if our current data-driven approach can reproduce this success.
Critical weaknesses
  • The presentation is hard to follow for me. It contains many forward references and unexplained notations. I struggled to understand the source of quite some equations
  • I am unsure about the main models being tested and how they connect back to successfully track back HH.
Questions for authors

Questions

  • What is z0z_0 in equation 3? The notations should be better explained.
  • I don't get where eq.4 is from. I thought, from Fig. 2 zz's are learned and there is no a priori assumptions about the temporal dynamics. Then why there is a fixed updating equation for the learned latent?
  • How much physical assumptions are put in the learned model?

Some requests

  • I think eq.1-2 can be explained better for a human reader. The notations should be explained by their physical meanings then interpreted as e.g., command.
  • The paragraph starting line 198 is confusing. Many of the words have specific meanings like gates/trajectories etc. but was not clear when the reader is here. This often happens when using AI tools iteratively as it sees the later part of the paper. But that is not the case for a reader yet.
Reviewer_Fk8BOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

This paper is a largely LLM-written (and probably conceived) paper about inferring neural firing dynamics.

Critical strengths

None.

Critical weaknesses

The paper basically says "you can predict how a neuron will respond without uncovering or inferring its actual mechanism."

The paper completely and entirely overlooks "Learning a synaptic learning rule" and "The Virtual Patch Clamp" and the whole huge literature of program synthesis work for mechanism discovery. In other words, the approach taken in this paper and the meagre results generated by it, are completely off-base and unrecoverable.

My guess is that the Claude prompt for this paper was something like "can we infer Hodgkin Huxley's rules from observed neural trace data" but prompted by someone who does not know the literature and set of available techniques well enough to correct the LLM away from dutifully saying that this is cool and novel.

Questions for authors

None. No need to engage. My view will not change.

Reviewer_6SQgOpenReview ↗
Rating
2
Confidence
2
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

The paper asks whether a model trained only on current and voltage from simulated Hodgkin–Huxley neurons recovers the hidden gating variables and the dynamics that govern them. The authors train a structured latent model with GRU encoders, latent variables bounded in (0,1) with voltage-dependent first-order relaxation, and a conductance-like current readout. The gates n, m, h are withheld during training and model selection. After the model is frozen, they fit a regression map (the "chart") from the latent state and voltage to the gates. They then evaluate four things: response prediction, gate decoding in distribution, gate decoding on unseen protocols, and whether the learned latent dynamics, pushed through the chart with the chain rule, match the HH gate equations.

The main findings: • Prediction error and seed-to-seed spread drop at K = 3. • OOD gate decoding keeps improving up to K = 5–6. • Under voltage clamp, m is decoded with R² > 0.99, but the transported dm/dt agrees poorly with HH (R² about 0.1–0.3). • Several baselines (S4D, GRU, Latent ODE) show the same gap. • A positive control using the true gates behind known invertible transformations passes the same audit, so the gap is not an artifact of the evaluation.

Critical strengths

• The question is a good one. Decoding physical variables from a latent space is often treated as evidence that a model has learned the mechanism, and this paper tests that assumption directly. • HH is a sensible testbed because the hidden states and equations are known. • The experiments are careful. The gate labels are kept out of training and selection, splits are by whole trajectory, chart regularization is chosen by grouped CV, and the OOD evaluation is done once. • The positive control for the audit is important and well done. • There are many robustness checks (chart class, chart capacity, time resolution, branch count, noise, baselines), and the authors are careful not to overclaim. • Figure 3 was helpful in showing how prediction accuracy and state recovery change with the number of latent dimensions.

Critical weaknesses
  1. The m-state R² is dominated by voltage. On the smooth voltage-clamp points used for this test, m reaches m∞(V) in around a millisecond following each command change, so voltage alone predicts m with R² = 0.976 (Table 2B). The headline R² > 0.99 is therefore not by itself strong evidence of state recovery. The interesting part is the gap m − m∞(V), which also determines dm/dt. The latent state captures much of it (the paper reports a 67–77% reduction in error over voltage alone), but dm/dt equals this gap divided by τm, so the remaining error is strongly amplified where τm is short. This is consistent with Table 10: where τm < 0.16 ms, the transported dm/dt is nearly uncorrelated with the true dm/dt, while state R² stays at 0.992. I suggest reporting R² on m − m∞(V), split by τm range, as the main state metric for m.

  2. No reference for how good the fit could be. HH is contained in the model class (K = 3, two branches with exponents (0,3,1) and (4,0,0), HH conductances and reversal potentials, and HH rate functions in the kinetics network). However, the paper does not report how well this HH setting performs under the same encoder and evaluation. The trained models reach NRMSE 0.45 at K = 3 and reproduce the spike count exactly in only about 32% of windows, while S4D reaches NRMSE 0.19 on the same data. Without a reference value, it is unclear whether the state–dynamics gap reflects a real limitation of what can be learned from current and voltage, or models that are far from the best achievable fit. Please report the performance of the model set to the true HH kinetics. Please also train from an HH initialization and check whether τm drifts away from HH while prediction stays equally good. That would directly show whether τm is identifiable from this data.

  3. Missing connection to classic HH reduction results. It is well established that m can be approximated by m∞(V) and that n and h are approximately linearly related, so HH can be reduced to a two-variable system of voltage and one recovery variable (Rinzel 1985; Kepler, Abbott & Marder 1992). Meunier (1992) further showed where the instantaneous-m approximation breaks down. This literature is not cited, but it is directly relevant: it explains why voltage alone predicts m so well, why gates are already decoded well at K = 1–2 (Appendix A.4), and why the remaining fast m dynamics are the hardest part to recover.

  4. The field metric is dominated by samples just after command changes. On constant-command segments, m reaches m∞(V) within about 1 ms, so the true dm/dt is near zero except right after a command change. In a simple simulation with 10-ms steps, the first 1ms after each step contains about 10% of samples but over 99% of the variance of dm/dt. The reported field R² therefore mainly reflects these few samples. The far-event analysis (>2 ms) is helpful, but please also report the field error separately for the first 0–2 ms after each command change, since this is where nearly all of the change in m happens. I also suggest extending the one-step comparison in Table 7 to longer time windows. Starting from the decoded gate values at a given time, integrate the HH equations forward for several milliseconds under the same voltage command, and compare this HH trajectory with the gate trajectory decoded from the model over the same period. If the learned dynamics are correct, the two trajectories should stay close; if not, this test shows how quickly they diverge.

Page 5: The transported field is important to the main argument, but I found the explanation a little difficult to follow. A simple numerical or visual example before Equation 6 might help readers understand it better. Page 9, Discussion: Since the authors use the true HH gating variables to evaluate state recovery, how could a similar evaluation be performed on real neuronal data where these hidden states are not directly available?

Questions for authors
  1. Clarity. • The terms "chart", introduction before they are defined. "transported field",and "generator" are used in the abstract and • Much of the key evidence is in the appendix (16 additional tables). • The protocol selection score (log of minimum eigenvalue plus 0.25 times model disagreement) is not explained, and its effect on the main result is not tested.

Other Questions

  1. What prediction error does the structured model get when set to the true HH kinetics?
  2. What is R² on m − m∞(V) for the charts in Table 2B, split by the τm quartiles of Table 10?
  3. How do the learned τ_k(V) compare with the HH time constants? Under voltage clamp these can be compared directly, without a chart.
  4. What exponents, conductances, and reversal potentials are learned? Does any branch look like Na or K? Does the exponent penalty push the model away from the HH powers?
  5. How much does prediction error change if τm is scaled by about 6×? Is τm practically identifiable from this loss and these protocols?
  6. Why is the Latent ODE only evaluated at K = 2?

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.