acceptodds
Withdrawn from ICLR 2027

Brain-Plausible Local Learning in SNNs: Training Hodgkin–Huxley Spiking Neural Networks with Predictive Coding

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Almost all deep spiking neural networks (SNNs) use Leaky Integrate-and-Fire (LIF) neurons and are trained by backpropagation through time (BPTT) with surrogate gradients. Both choices are made for tractability, not biology: LIF throws away the ion-channel dynamics that generate real action potentials, and BPTT needs a global error signal no cortical mechanism is known to implement. Conductance-based models such as Hodgkin–Huxley (HH) are avoided because unrolling their coupled nonlinear dynamics through time is expensive and numerically fragile. We show the trade-off is avoidable. Our method trains a network of HH neurons for classification with no backpropagation, no surrogate gradients, and no unrolling, using predictive coding (PC) to supply layer-local errors. One design choice makes this work: a rate proxy, the spike count of each HH population divided by the simulation window, which gives PC the continuous states it needs. No gradient ever crosses spike generation, so the cost of learning does not depend on how complex the neuron is. Across MNIST, Fashion-MNIST, N-MNIST and the Spiking Heidelberg Digits, HH matches a LIF baseline trained under the same objective while firing fewer spikes in 26 of 28 window-matched comparisons. Our learning interface allows biophysically realistic neurons to retain competitive accuracy.

Primary area:
unsupervised, self-supervised, semi-supervised, and supervised representation learning
Keywords:
spiking neural networks, predictive coding, Hodgkin-Huxley neurons, local learning, biologically plausible learning, spike efficiency

Reviews

Among:
Shape:
1234this paper, Bottom 2%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.33 over 3 reviews, above 2% of the others.
Reviewer_DvN8OpenReview ↗
Rating
2
Confidence
3
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

This paper proposes a PCHH-SNN framework to study Spiking Neural Networks (SNNs) with Hodgkin-Huxley (HH) models and a local learning rule. Specifically, the authors first model the SNN based on HH neurons. Then, to balance the complexity of HH dynamics and the efficiency of SNN training, the authors introduce a local learning rule (local PC) to construct a gradient-free training process. The experimental results on MNIST, Fashion-MNIST, SHD, and Neuromorphic MNIST demonstrate the effectiveness of the proposed method.

Critical strengths
  • Compared to Leaky Integrate-and-Fire (LIF) neurons, the HH model is more biologically plausible for SNN modeling.

  • Compared to existing BP-based SNN training methods, the proposed PCHH-SNN trains the model without gradients, avoiding the gradient mismatch problem associated with surrogate BP and improving training efficiency.

Critical weaknesses
  • The datasets and model architectures used in the experiments are relatively simple. The effectiveness of PCHH-SNN on more complex datasets (such as CIFAR-10/100 and CIFAR10-DVS) and network architectures (such as VGG, ResNet, and ViT) needs further validation.

  • In addition to the spike rates reported in Tab. 3, there is a lack of comparison between HH-based SNNs and LIF-based SNNs in terms of inference energy consumption. Moreover, a comparison of training efficiency between PCHH and BPTT-based training is not provided in the paper.

  • Although the HH model is more biologically plausible than the commonly used LIF model, its dynamics are also more complex, reducing the computational efficiency of SNNs. In addition, the results in Tab. 1 do not show that HH outperforms LIF/IF across all tasks. The benefits of the HH model for SNNs need further discussion.

Questions for authors

Please refer to "Critical Weaknesses".

Reviewer_8J6cOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
3
Soundness
1
Read review
Paper summary

The authors propose HH model-based SNNs that can be trained by bypassing the complex dynamics of HH models. The bypass is realized by “rate proxy” that is nothing but layerwise average spike firing rate, which is used as a continuous layer feature for each neuron. This continuous state replica of the HH-based SNN is trained using the predictive coding scheme. The feasibility of this framework was identified on four simple datasets (MNIST, F-MNIST, N-MNIST, and SHD). The main advantage of this HH-based SNN over the conventional LIF may lie in its spike economy.

Critical strengths

none

Critical weaknesses
  1. Limited novelty: There exist several attempts to overcome the difficulty in SNN training by decoupling the forward pass and learning with predictive coding schemes. The rate proxy (considered as one of the main contributions) is merely the spike firing rate average that ignores local rate changes over time. This average rate is taken as a continuous state for each neuron in each layer. In this regard, the novelty of this manuscript is marginal.
  2. Weak objective of HH usage: The learning is decoupled with the forward pass involving the complex dynamics of the HH model. Thus, I am very puzzled about the objective of HH usage, and thus the advantages of HH over the conventional LIF or IF. The only claim lies in “spike economy”. I do not think the spike economy is an inherent characteristic of conductance-based models with high complexity like HH. The same is likely for LIF with tuned hyperparameters.
Questions for authors

Have the authors attempted to optimize the hyperparameters of the LIF models used in the baseline to enhance its spike economy?

Reviewer_sARyOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
3
Soundness
1
Read review
Paper summary

The paper separates a forward spiking simulation from predictive-coding inference. Spike counts divided by the simulation window initialize continuous PC states; these states are then optimized and used for local weight updates without differentiating through spike generation or the neuronal ODEs. The method is evaluated against LIF-PC on four classification datasets, with a five-seed comparison, spike-count sweeps and additional ablations. The main claim is that HH neurons can be trained without BPTT while retaining accuracy and reducing spikes.

Critical strengths
  1. An explicit learning procedure with relevant controls

    The paper gives an explicit learning procedure, separates the spiking and continuous inference phases, and reports neuron and preprocessing settings in considerable detail.

    The paper includes analog and encoding controls and an SHD comparison with and without PC settling. It also acknowledges that SHD is represented as a static flattened frame and that fewer spikes do not establish lower digital simulation cost.

  2. Transparent reporting of repeated-run results

    The five-seed HH/LIF comparison includes per-seed results, a failed HH run, uncertainty and multiple-comparison caveats. The authors appropriately avoid claiming a significant accuracy advantage from the Fashion-MNIST p-value alone.

Critical weaknesses
  1. The implemented neuron does not establish the claimed HH interpretation

    The implemented neuron does not straightforwardly support the claim of full HH action-potential dynamics. Section 3.3 emits a spike at 1.2 mV above rest (0.8 mV on SHD), resets voltage to zero, and imposes an external refractory counter. It also applies V = 30 tanh(V/30) after every Euler step. These are substantive modifications, rather than merely detecting action potentials produced by the HH dynamics. Since the main motivation explicitly contrasts HH ionic spike generation with LIF threshold/reset, evidence is needed that the advertised biophysical mechanism survives this implementation and contributes to the results.

  2. The link to the spiking computation and the accuracy–spike claims is unresolved

    The continuous PC objective is not shown to train the actual spiking computation in the claimed sense. Equations (3)–(5) use clipped affine predictions, while the actual spike-rate response is generated by a different, fixed HH current-to-spike process. The HH rates initialize PC states but are not constrained to remain consistent with them; x(0) is supplied directly from the image or averaged events. At a converged test-time fixed point of the stated updates, the free output has zero prediction error and the hidden state equals its prediction from x(0), giving the corresponding continuous feedforward classifier independently of the initial HH rates. Finite settling can retain an initialization effect, but that is different from demonstrating a consistent learned HH input–output model. The analog control in Table 4 and the PC-versus-FF comparison in Appendix H are useful, yet do not directly separate this initialization effect from neuronal computational value.

    The accuracy and spike conclusions need narrower statistical and mechanistic scope. Failure to reject an accuracy difference with five seeds is not an equivalence or non-inferiority test. For N-MNIST, Table 2 gives 91.58% for HH versus 93.45% for LIF, with much larger HH variance. The failed seed is part of the method's reliability. For spike counts, Table 3 is a single run per cell, and the imposed 2 ms refractory period exceeds the vision simulation windows up to T=60. Appendix H itself shows that changing the SHD refractory period reduces HH spikes about 8.5-fold with almost unchanged accuracy. Thus, the evidence does not yet distinguish an intrinsic HH advantage from effects of operating point, refractory settings and continuous PC inference.

    Decoupling PC from neuronal simulation is potentially useful, but the distinctive role of biophysical HH dynamics has not been established.

    Avoiding BPTT through HH does not remove the cost of forward HH simulation or 100/50 continuous PC iterations during training/testing. Likewise, the local outer-product update should be distinguished from the full implementation, which also uses shared feedback weights, Adam and global gradient-norm clipping. These qualifications would improve the biological and computational interpretation.

Questions for authors
  1. Can you provide representative voltage/gating traces or existing ablations showing which HH action-potential properties are retained under the low threshold, hard reset and per-step tanh saturation? What evidence supports calling this implementation full HH rather than a modified conductance-based threshold/reset model?
  2. Please explain the intended consistency between the PC prediction function and the HH rate response. For fixed learned weights, how do prediction accuracy and final PC states change when HH proxy initialization is replaced by zero, random or analytic feedforward initialization, and as the number of inference iterations increases? How does this relate to the existing analog and FF controls?

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

Related papers

Discussion (0)

Sign in to comment.