acceptodds
Withdrawn from ICLR 2027

Time Will Give You the Answer: Controllable Factorial Permutations in Spiking Neural Networks

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

The brain works through spiking neural networks. In this paper, we investigate the relative order of the first spikes across neurons. The relative order of first spikes across neurons creates a rapidly growing discrete state space. The full first-spike order of neurons can represent discrete states. For , this gives 3.6 million possible orders, and the number grows factorially with . We show the ordering space is controllable and trainable. By updating input current or through intrinsic parameters alone—membrane time constant, firing threshold, and membrane resistance—we can reach all permutations. We also design an optimizer that moves the neurons into any target order by updating each neuron's membrane time constant. We picked a 30-class dataset. In a 6-spiking-neuron system, we assign each dataset class an ordered pair of neurons. For example, means we want neuron A to fire first and neuron B to fire second; neurons firing after the second are ignored. We reach an accuracy of 27.7% 1.03 across 5 different seeds, compared to 3.33% random chance.

Primary area:
applications to neuroscience & cognitive science
Keywords:
Spiking Neural Network

Reviews

Among:
Shape:
1234this paper, Bottom 1%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.00 over 4 reviews, above 0% of the others.
Reviewer_imCSOpenReview ↗
Rating
1
Confidence
4
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

The paper treats the relative order of first-spike times as a representation in spiking neural networks. For a population of MM neurons, the full firing order can in principle encode M!M! permutations, and an ordered prefix of the first kk neurons gives P(M,k)=M!/(M−k)!P(M,k)=M!/(M-k)! possible codes.

The authors first derive the analytical first-spike time of an LIF neuron. They then ask whether any target firing order can be reached by adjusting the input current or the neurons' intrinsic parameters. With six neurons, they report that all 720 permutations, and all pairwise transitions between them, are reachable. They also propose a simple target-aware optimizer that adjusts membrane time constants to produce a desired order, and apply it to populations of seven and eight neurons without retraining its two constants.

The second half looks at firing-order representations for classification. The authors start with the identity of the first neuron to classify six-class synthetic data and MotionSense. They then use the ordered identities of the first two neurons in a six-neuron population to represent 30 classes from an EMNIST subset. The best class-to-order assignment, built from training-set class centroids, reaches 27.7%27.7\% accuracy. The paper also reports several negative results and discusses the limits of the approach.

Critical strengths

1. The paper looks at an interesting and fairly underexplored representation: the identities and relative order of several first spikes.

Going beyond a single first-spike decision to ordered prefixes gives a natural combinatorial code space. The paper separates first-spike coding from top-kk ordered firing identities cleanly, and its controlled experiments do illustrate some properties of the representation.

2. The authors are refreshingly open about negative results and limitations.

They report failed experiments, preregistered criteria that were not met, the weaknesses of the supervised class-to-order construction, and a correction to an earlier implementation. That openness helps the reader judge what the experiments do and don't establish.

Critical weaknesses

1. The central controllability result tells us much less than the "factorial state space" framing suggests.

The analytical first-spike time is ti∗=−τilog⁡(1−Vth,iRiIi)t_i^{\ast}=-\tau_i\log\left(1-\frac{V_{\mathrm{th},i}}{R_iI_i}\right). In the setting used by the order optimizer, RiIi=2Vth,iR_iI_i=2V_{\mathrm{th},i}, so ti∗=τiln⁡2t_i^{\ast}=\tau_i\ln 2.

Getting an arbitrary firing order by independently changing the membrane time constants then boils down to choosing a set of continuous scalars whose ranking matches the target permutation. Given this, the fact that all M!M! permutations are reachable is close to what the formula already tells you. Exhaustively checking every permutation for six, seven, and eight neurons confirms that the implementation works as intended, but it doesn't show any non-trivial learning or representational property of an interacting SNN.

This takes a lot away from what I took to be one of the paper's main technical contributions.

2. More importantly, the main classification experiments don't convincingly show a benefit from spiking dynamics or from the factorial state space.

The paper says that all neurons in the classification model share τ\tau, RR, and VthV_{\mathrm{th}}. Under exactly that condition, the paper's own derivation gives argsort⁡(t∗)=argsort⁡(−I)\operatorname{argsort}(t^{\ast})=\operatorname{argsort}(-I).

So in the main classification setting, the relative firing order is just the ranking of the learned input currents. You could describe the model as mapping the input to learned currents and then reading off the indices of the top-ranked outputs, with the LIF first-spike dynamics supplying only a monotonic transformation that leaves the ordering untouched.

That raises a basic question: what do the spiking dynamics actually contribute to the reported results? The fact that a factorial number of permutations exists mathematically doesn't mean a learning system can make good use of them. The experiments show as much. The best 30-class EMNIST result is only 27.7%27.7\%, against 73.1%73.1\% for the reported linear model, and that best result depends on a supervised, geometry-aligned class-to-pair assignment built from labeled training-set class centroids. The paper also says its attempt to find a useful factorial structure without such a label-derived code did not work.

All in all, I think studying multi-spike order is an interesting idea, but the current results don't show that the proposed factorial representation is a substantive learning mechanism or offers enough of an advantage to meet the bar for an ICLR contribution.

Questions for authors

Q1. Beyond ranking independently adjustable continuous variables, what non-trivial property do the factorial controllability experiments establish?

In the optimizer experiment, ti∗=τiln⁡2t_i^{\ast}=\tau_i\ln 2. If each τi\tau_i can be set independently over a wide enough interval, it seems that any permutation follows directly by assigning the target ordering to the τi\tau_i. Could the authors clarify which part of the exhaustive controllability result would not follow immediately from this monotonic relationship?

Q2. When all neurons share τ\tau, RR, and VthV_{\mathrm{th}}, what role do the LIF dynamics play in the main classification results?

In this setting, the manuscript derives argsort⁡(t∗)=argsort⁡(−I)\operatorname{argsort}(t^{\ast})=\operatorname{argsort}(-I). Therefore, in terms of predicted ordering, the classifier appears equivalent to directly ranking the learned input currents. Is there any classification behavior in the reported experiments that cannot be reproduced by removing the LIF first-spike transformation and using the current rankings directly?

Q3. Can the authors show that the factorial firing-order representation is useful beyond a hand-designed class-to-code mapping?

The strongest 30-class result relies on a geometry-aligned assignment constructed from labeled training-set class centroids, while the label-free factorial self-organization experiment is reported as unsuccessful. The factorial representation also performs substantially worse than the reported linear baseline. Could the authors provide stronger evidence that useful factorial structure can emerge through learning itself, rather than being largely supplied by the class-to-order assignment?

These questions concern the central claims of the paper rather than suggestions for future work. My evaluation would depend on convincing evidence that the proposed representation exploits non-trivial spiking dynamics and provides functionality beyond ranking a small number of learned scalar outputs.

Reviewer_5M94OpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
1
Soundness
1
Read review
Paper summary

The paper uses time ordering coding of spiking neurons for classification tasks. The neurons are not connected, but interrelated through the ordering of their spike times. Classes and their encoding are specified by hand. Loss functions ensure that the desired spike order is achieved for each class and, additionally, that the task-relevant spikes are separated by temporal intervals of some desired length. Learning takes place through membrane time constant or input weight parameters and biases.

Critical strengths

It is interesting to consider spike order coding.

Critical weaknesses

The extremely short literature review does not cover the current state of the art on learning of spiking systems. For example, many papers have used time to first spike coding, not only Comsa et al. 2020, many papers have investigated single spiking neuron learning. Many statements are trivial or well known. The paper is not up to the state of the art on spiking neural systems learning. The tasks are not well described.

Questions for authors

How does the performance compare to training ReLU networks?

Reviewer_r9duOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

Weaknesses

  1. Poor abstract. Informal wording, such as “crazy density,” and vague dataset descriptions obscure the contribution and scientific context.

  2. Unclear motivation and novelty. The introduction does not clearly establish a research gap beyond existing rank-order coding.

  3. Weak presentation. The method lacks an overview figure explaining input encoding, spike-order control, and classification.

  4. Insufficient methodological justification. The classification model does not use the proposed control optimizer, and its shared neuron parameters make spike ordering equivalent to simply ranking the input currents.

  5. Weak results. The 27.7% accuracy is much lower than the linear baseline’s 73.1%, without establishing a convincing accuracy–parameter trade-off, as displayed in Section 7.

Conclusion

Unclear novelty, weak methodological justification, unconvincing results, and incomplete documentation prevent acceptance in its current form.

Critical strengths

None

Critical weaknesses
  1. The methodological contribution is insufficiently validated. With shared neuron parameters, classification reduces to ranking input currents and does not use the proposed control optimizer (Sections 3 and 7). Consequently, the classification results do not establish a clear learning benefit from the paper’s central control mechanism.

  2. The experimental evidence does not establish a convincing benefit. The main accuracy is 27.7%, versus 73.1% for a linear baseline (Section 7). Although the proposed model uses fewer parameters, no comparison at a similar parameter budget establishes a worthwhile trade-off.

Questions for authors

None

Reviewer_nTXqOpenReview ↗
Rating
1
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper studies first-spike rank codes for constant-current LIF neurons. It tests control of firing permutations through neuronal parameters, proposes a target-aware time-constant update, and uses the first two neurons of a six-neuron population as a 30-class output code. A training-label-derived centroid assignment improves EMNIST accuracy over lexicographic and random assignments, with additional first-spike classification and negative-result diagnostics.

Critical strengths
  1. The class-to-code assignment experiment isolates a concrete representational choice: changing the supervised assignment raises top-two accuracy from 11.878% to 27.682%, with five-seed results and a rank-wise error decomposition. The distinction between available codewords and successfully learned use of those codewords is useful.

  2. The paper explicitly reports its target-aware control setting, the use of training labels for code construction, the reduction to current ranking when classification parameters are shared, and several unsuccessful experiments. These qualifications help bound the interpretation of the results.

Critical weaknesses
  1. The central controllability and optimizer results have an elementary direct construction that is not addressed as a baseline. In Section 3.4, t_i=ln(2) tau_i. Any feasible target order with margin delta can therefore be realized by directly setting tau_(pi_k)=tau_min+(k-1)delta/ln(2), subject to the stated upper bound. There is no need to learn an iterative optimizer for the demonstrated uncoupled setting. Furthermore, differentiating the softplus adjacent-gap objective at temperature T_opt gives dL/dtau_(pi_k)=ln(2)/(M-1) times (q_k-q_(k-1)); Equation 7 is projected gradient descent on that convex objective, with constants absorbed into the step size. Exhaustive success at M=6,7,8 consequently does not by itself establish a new optimization or controllability result. The recurrent-coupling statement could be more interesting, but Appendix C does not specify that coupled system and its optimization protocol sufficiently to establish what goes beyond the direct case.

  2. The classification evidence does not establish a computational advantage specific to spiking dynamics. Section 7 explicitly states that all classification neurons share tau, R, and threshold. Equation 1 and the monotone current map then make the predicted ordered pair exactly the top-two ranking of six affine scores. A non-spiking six-score rank classifier can produce the same decisions. The paper does report a larger 30-output linear baseline at 73.1%, which I acknowledge; the missing comparison is the identical six-score, identical-parameter rank model and an explanation of what novel learning insight remains after this equivalence. Beating chance or the 20% ceiling of a restricted first-identity decoder establishes that a larger output code is used, but does not establish an advantage over that matched rank-code alternative. This is a contribution/significance concern, rather than a rejection based solely on the low absolute accuracy.

The recommendation is based on the limited advance once the direct-control construction, projected-gradient interpretation, and exact classification equivalence are accounted for. The negative experiments and modest accuracy are not, by themselves, grounds for rejection.

Questions for authors
  1. What result remains beyond direct feasible time-constant assignment and projected gradient descent in the uncoupled control setting? For the coupled setting, please give the actual recurrent equations, parameter constraints, success criterion, and a comparison that isolates the nontrivial control problem.

  2. Since the shared-parameter classification rule is exactly a ranking of six affine scores, what distinction would remain against a six-score non-spiking rank classifier trained with the same pair-margin objective and code assignment? Please identify the substantive new representational or learning result beyond the known combinatorics of rank-order coding.

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

Related papers

Discussion (0)

Sign in to comment.