acceptodds
Under review as a conference paper at ICLR 2027

Multi-Rate Dendritic Memory Spiking Neural Networks for Multi-Timescale Neural Signal Decoding

Abstract

Invasive neural recordings capture signals with distinct timescales, including action potentials, local field potentials, and dopamine concentrations. Decoding behavior from these multimodal streams requires combining signals with different timescales, sampling rates, and temporal information densities while preserving modality-specific discriminative information. Existing methods often struggle to explicitly and independently control the temporal resolution at which history is updated and the span of history that is retained, making this integration difficult. Inspired by the proximal-to-distal functional gradient of biological dendrites, we introduce MRD-SNN, a spiking network in which each neuron maintains a direct somatic pathway and multiple sparse dendritic memory branches. Each branch refreshes at its own rate and uses a Legendre representation to summarize history over a specified nominal physical horizon. We evaluate the model on three cross-modal neural decoding tasks. MRD-SNN achieves the highest accuracy on all three tasks. Ablations on multimodal neural recordings and a spiking speech benchmark show that temporal resolution and memory horizon jointly determine what a branch can capture: the update interval controls the temporal granularity of input aggregation, while the memory parameter controls the nominal span of represented history. In multimodal neural decoding, branches need sufficient coverage to capture discriminative trends spanning several seconds. In tasks that depend on local timing patterns, finer resolution preserves transient structure. Branch-combination ablations further show that, when a task requires both fine temporal detail and longer context, multiple update rates can provide complementary information and outperform the tested single-scale configurations.

Primary area:
unsupervised, self-supervised, semi-supervised, and supervised representation learning
Keywords:
spiking neural networks, dendritic computation, multimodal neural decoding, multi-timescale signal processing

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_8UuEOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes MRD-SNN, which augments spiking neurons with dendritic memory branches operating at different update rates and memory horizons. Each branch maintains a Legendre memory and supplies a cached readout between updates. The method is evaluated on three neural decoding tasks, SHD, and sequential MNIST. The authors also study memory capacity, branch combinations, and the accuracy-efficiency trade-off of reducing memory update frequency.

Critical strengths
  1. The paper provides a concrete way to combine temporal memory with sparse updates. Separating the memory horizon from its update frequency is useful for controlling computational cost. The matched-horizon comparison reports 1.55-1.78x speedups with accuracy differences of at most 0.74 percentage points, which supports the practical value of reducing refresh frequency.

  2. The empirical results are promising.

Critical weaknesses
  1. It is not convinced that the accuracy gains come specifically from using different update rates. In Table 4, changing the update interval at fixed θ\theta also changes the physical memory horizon. When the horizons are matched in Table 9, full-rate and multi-rate updates achieve similar accuracy. These results make a stronger case for reducing computation than for improving accuracy through rate diversity.

  2. The shared-versus-independent memory comparison in Table 5 also changes the memory budget substantially: moving from G=1G=1 to G=128G=128 increases the number of dendritic memory-state variables 128-fold. It is therefore difficult to tell how much of the improvement comes from giving each neuron its own memory and how much comes from simply having more memory capacity.

Questions for authors
  1. Could you repeat the mixed-rate versus same-rate comparison on SHD and sMNIST while keeping the same set of physical memory horizons? Does the advantage reported in Table 4 remain under this control?

  2. Do neuron-specific memories still outperform shared memories when the total number of memory-state variables is matched?

1 comment

Q1:

We thank the reviewer for suggesting this important control. We added comparisons that keep the same pair of physical memory horizons while changing only the refresh intervals, allowing us to examine the effect of update scheduling separately.

The horizons are 0.112 and 1.12 seconds on SHD, and 98 and 980 input steps on sMNIST. Each branch has q=10q=10; parameter counts, connection masks and trainable initializations are identical. We adjust θ\theta with nn to keep H=nθΔtH=n\theta\Delta t fixed.

Refresh scheduleIntervals nSHD accuracy (%)sMNIST accuracy (%)
Mixed[2,20]82.77 ± 0.3993.50 ± 0.14
Both high rate[2,2]83.49 ± 0.8393.94 ± 0.21

Results use five folds and three seeds per fold, with validation-selected checkpoints. Mixed and high-rate refresh achieve nearly the same accuracy: mixed refresh is lower by approximately 0.7 and 0.4 percentage points, while reducing nominal branch-update counts by approximately 45%.

Our design considers both memory specialization and computational cost. A high-rate branch can obtain a long horizon through a larger θ\theta, but frequent refresh does not necessarily imply a rapid response to new input. At fixed memory order, when ∥A0∥/θ≪1\lVert A_0\rVert/\theta\ll 1, the branch update can be approximated as

mj≈mj−1+1θ(A0mj−1+B0xj)m_j\approx m_{j-1}+\frac{1}{\theta}(A_0m_{j-1}+B_0x_j)

Here, mjm_j is the memory state, xjx_j is the current input, and A0A_0 and B0B_0 are the unscaled Legendre matrices. A larger θ\theta weakens both the write from new input and the change to the previous state at each update, making the branch a slowly evolving summary of history. A high-rate, long-horizon memory therefore does not automatically provide both recent detail and distant context.

Appendix G further illustrates the risk of performance failure for a single high-rate, long-horizon branch on tasks requiring very long retention together with recent timing information. On the 8192-step XOR task, increasing the single branch to q=20q=20 yields only 57.75 ± 12.21% accuracy, whereas a high-rate short-memory branch paired with a low-rate long-memory branch reaches 99.98 ± 0.03%.

We therefore let different branches specialize in recent detail and distant context, and reduce the long branch’s refresh rate to lower update costs. When the required memory span is shorter, two high-rate branches with different θ\theta values also perform well; multirate scheduling can retain similar accuracy while reducing update overhead.

Q2:

We appreciate the reviewer’s observation that the original comparison also changes memory capacity. We therefore added comparisons on all three neural decoding datasets that keep the total number of dendritic memory states fixed while varying the degree of memory sharing.

We retain the four-branch architecture, update intervals, nominal memory horizons and training protocol of Table 5, and adjust the memory order qq as the number of memory groups GG changes. Every configuration has the same total number of dendritic memory states:

S=2×4×Gq=10,240,Gq=1,280S=2\times4\times Gq=10{,}240,\qquad Gq=1{,}280

Here, G=1G=1 shares memory across the entire layer, whereas G=128G=128 gives each neuron its own memory.

GqSpike–LFP–DADANDI 000121DANDI 000447
1128065.67 ± 3.7879.84 ± 2.1362.84 ± 2.52
816073.56 ± 5.0981.59 ± 2.9961.92 ± 0.73
324074.12 ± 6.2083.62 ± 1.4770.08 ± 3.46
642074.00 ± 5.6585.47 ± 0.8871.41 ± 4.75
1281075.20 ± 5.6886.02 ± 1.7975.42 ± 2.46

Accuracy (%) is reported as the mean ± sample SD across five folds, after averaging three seeds within each fold.

Neuron-specific memory achieves the highest mean accuracy on all three datasets under the matched state budget, exceeding the best shared-memory configuration by 1.08, 0.55, and 4.01 percentage points, respectively. These results support the value of maintaining separate temporal contexts for individual neurons even when the total number of memory states is fixed.

Reviewer_f75xOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes a Multi-Rate Dendritic Memory Spiking Neural Network (MRD-SNN), which decodes neural signals containing information at different temporal scales by means of dendritic branches before passing through an LIF neuron. In the proposed network, each neuron has a direct somatic pathway and K sparse dendritic memory branches; each branch has its own update interval, Legendre state, and physical memory horizon, and updates at different rates. MRD-SNN performs better than selected baselines on the three neural decoding tasks for cross-modal neuronal signals such as action potentials, LFP, and dopamine.

Critical strengths
  1. Proposed architechture independently controls how long information is retained(Hk=nkΔθkH_k = n_k\Delta\theta_k) and how frequently each branch should update(nkn_k)
  2. Diverse ablations: Ablation studies also suggest that longer memory horizons and complementary fast and slow branches can improve performance. The paper further reports a reduced memory update frequency and improved runtime in a streaming implementation.
  3. Results reported in Table 1 are consistently strong across three neural decoding benchmarks, compared against baselines with more than one type of temporal model. The proposed method is competitive with models incorporating adaptive dynamics such as ALIF and explicit memory: LMU and GRU.
Critical weaknesses
  1. The multi-rate mechanism is not convincingly validated in the paper. Ablations are very limited.

  2. The performance for SHD and sMNIST is not compared with other methods. Table 4 shows results on the dual fast+slow method(as I understood it: MRD-SNN); a comparison with other models in the benchmark should have been provided. Moreover, SHD and sMNIST are too simple datasets to understand the benefits.

  3. Given that the central contribution of the paper is to generalise to multi-scale cross-modal tasks, Table 1 benchmark only evaluates the neural-decoding tasks, and Table 4 shows evaluation for sequential classification only, showing the fast vs. slow branch ablation, so the results provided are insufficient to claim generalizability. Evaluation should include diverse tasks and datasets (LRA, PSMNIST, SSC) for generality.

  4. In Table 9, the authors report fewer memory updates (speedup), but it is measured using an uncompiled PyTorch implementation rather than neuromorphic hardware deployment. Thus, to claim efficient deployment would benefit from more complete accounting of total computation and memory cost.

Questions for authors
  1. In Section 4.1, the authors evaluate the method using three datasets differing in recording source, decoding task, number of trials, available signal modalities and sampling rates. How do the authors distinguish the effect of temporal diversity from other dataset-specific factors such as Δ\Delta when improving performance?

  2. If possible, can the authors provide evidence of generalisation across subjects or recording sessions? Or evaluate the method on additional subjects?

  3. In Section 5.3, Table 4: the results justified for sMNIST are unclear. Why is the single slow branch outperforming the single fast branch (sMNIST accuracy: 84.93%, 54.11%), whereas for SHD the single slow branch is still performing well (SHD accuracy: 78.76%, 76.46%)?

  4. What are the energy measurements and latency in addition to memory consumption?

  5. How will your model perform with respect to recent spiking SSM models such as PSpikeSSM and SpikingSSM?

  6. How competitive is your model, MRD-SNN, on more complex tasks such as the Long Range Arena (LRA) dataset, SSC, etc.?

5 comments

Q1:

We thank the reviewer for pointing out the need to distinguish temporal diversity from other dataset-specific factors. We address this concern through comparisons within each dataset and clarify that the cross-dataset performance ordering is a descriptive observation rather than causal evidence.

To examine the role of temporal scales, we conducted comparisons within each dataset, keeping the task, data splits and preprocessing unchanged. The table below compares a single memory with eight branches spanning different memory horizons. All memories update at every step, and both models have 20,480 memory-state variables.

ConfigurationSpike–LFP–DA(%)DANDI 000121(%)DANDI 000447(%)
Single memory74.12 ± 4.5283.72 ± 1.8373.53 ± 3.02
Eight branches, eight horizons77.59 ± 4.4287.08 ± 1.4878.05 ± 1.62

Under the same memory-state budget, branches spanning multiple horizons achieve higher accuracy on all three tasks. Table 2 also shows that extending the memory horizon at a fixed update interval improves accuracy by 8.46, 3.67 and 0.64 percentage points, respectively. These within-dataset results show that branch configuration and memory coverage affect decoding performance, with benefits that depend on the task; however, they do not rigorously establish that cross-dataset performance differences can be explained solely by temporal diversity.

Q2:

Thank you for the suggestion. We added cross-session and cross-subject evaluations on DANDI 000121, comparing MRD-SNN with FSNN, ALIF and TC-LIF.

All models use identical splits, with three seeds per split.

Cross-session evaluation is performed within each animal: one complete recording session is held out for testing, another is used for validation, and the remaining sessions are used for training. The test session is excluded from both training and model selection. The two animals provide eight held-out sessions in total.

ModelCross-session accuracy (%)
FSNN66.35 ± 11.27
ALIF64.69 ± 7.58
TC-LIF66.02 ± 10.94
MRD-SNN70.68 ± 15.04

MRD-SNN achieves the highest mean accuracy, exceeding the strongest control, FSNN, by 4.33 percentage points, demonstrating transfer to unseen recording sessions from the same animal. However, performance varies substantially across sessions, indicating uneven cross-session generalization.

For cross-subject evaluation, models are trained and validated on one animal and tested on the other, in both directions. MRD-SNN achieves 50.07% and 48.88% accuracy in the two directions, while the other models also remain near the 50% chance level. Thus, the current results support within-animal cross-session generalization, but not direct cross-animal transfer. In invasive recordings, neurons and input channels do not have a natural correspondence across animals; cross-animal decoding in this setting requires further investigation of channel alignment or domain adaptation.

Q3:

Thank you for asking us to clarify this. The difference is related to the tasks’ memory requirements and prediction readouts, as explained below.

sMNIST presents an image one pixel at a time over 784 steps and predicts the class only at the final step. The short-memory branch has a nominal horizon of about 98 steps, whereas the long-memory branch spans about 980 steps. At the final step, the short-memory branch has difficulty retaining information from most of the earlier pixels, while the long-memory branch can cover the entire image. The single slow branch therefore achieves 84.93% accuracy, compared with 54.11% for the single fast branch.

SHD presents spike activity from multiple auditory channels at each time step. The short- and long-memory branches have nominal horizons of 8 and 80 time bins, respectively. Unlike sMNIST, SHD aggregates outputs over all valid time steps: the model can extract discriminative information as it hears a segment of the utterance and contribute that information to the final prediction, without having to reconstruct the entire utterance from memory at the end. Compared with the final-step readout in sMNIST, this readout reduces the need for a single branch to retain the entire history, allowing a shorter memory to be useful. Faster updates help preserve fine changes in spike activity within a segment, whereas slower updates aggregate neighboring time bins and can blur these differences. The single fast branch therefore achieves 78.82%, compared with 76.46% for the single slow branch.

Our design allows different branches to focus on local detail and longer history, accommodating task-specific requirements for temporal resolution and memory span.

Q4:

Thank you for the suggestion. We provide latency, energy and dynamic-state memory results below.

We tested the K=8 model on a single idle RTX 5090 with batch size 1, across five folds for each dataset. Both schedules use identical memory horizons and require 97 KiB of dynamic-state memory per sample. Multi-rate scheduling reduces updates but not stored states. Each table entry reports “multi-rate updating / every-step updating.” Measured GPU energy is obtained by integrating board power after subtracting idle power; estimated arithmetic energy is calculated from operation counts and reference per-operation energy costs.

DatasetLatency (ms)Measured GPU energy (J/trial)Estimated arithmetic energy (µJ/trial)
Spike–LFP–DA1206.8 / 2133.060.9 / 106.7195 / 1,178
DANDI 000121130.2 / 229.56.6 / 11.622 / 128
DANDI 000447421.2 / 746.621.3 / 37.672 / 413

Multi-rate updating achieves a 1.76–1.78× speedup, reduces measured GPU energy by approximately 43% and estimated arithmetic energy by approximately 83%, with consistent directions across all folds. On Spike–LFP–DA, each 10-ms time step requires approximately 1.2 ms of computation on average. At identical memory horizons, the accuracy differences between multi-rate and every-step updating are −0.60, −1.06 and +0.50 percentage points on the three datasets, respectively. These results show that multi-rate scheduling reduces computation time and energy while maintaining comparable accuracy.

Q5:

Thank you for recommending these baselines. We added P-SpikeSSM and SpikingSSM using the authors’ code, with the same inputs and splits and approximately matched parameter budgets.

ModelSpike–LFP–DA(%)DANDI 000121(%)DANDI 000447(%)
P-SpikeSSM77.55 ± 4.5085.86 ± 1.5175.15 ± 3.27
SpikingSSM77.43 ± 5.8287.79 ± 1.9478.65 ± 4.06
MRD77.15 ± 5.0487.19 ± 1.0478.42 ± 3.86

With the same inputs, splits, and approximately matched parameter budgets, MRD and SpikingSSM achieve similar mean accuracies: MRD is lower by 0.28, 0.60, and 0.23 percentage points on the three tasks, respectively. On the two DANDI tasks, MRD exceeds P-SpikeSSM by 1.33 and 3.27 percentage points. These results show that MRD is competitive on the evaluated neural decoding tasks.

Q6:

Thank you for the suggestion. We added experiments on SSC, a larger event-based speech benchmark with 35 classes and an official data split.

All SSC models use hidden width 128 and train for 100 epochs. Results report mean ± sample standard deviation across three seeds.

ModelSSC accuracy (%)
MRD (K = 8)73.45 ± 0.38
LMU69.33 ± 0.47
DH-LIF67.05 ± 0.16
TC-LIF62.81 ± 0.85
FSNN45.57 ± 0.25

MRD-SNN achieves the highest accuracy among these spiking and Legendre-memory baselines, exceeding the runner-up, LMU, by 4.12 percentage points. This result shows that its advantage also extends to a larger event-based speech task.

LRA includes text, image and path-finding tasks, but does not contain neural recordings or native multi-rate signals. We therefore prioritized SSC, which is closer to our application setting, to evaluate the model on a larger event-based benchmark.

Reviewer_f75xOpenReview ↗

Thanks - I would like more clarifications though

I am not sure why we should go with MRD instead of picking an already existing SNN like P-SpikeSSM.

Could you please shed some light on this with some empirical evidence? The MRD performance closely matches the existing ones.

  1. Please report LRA performance so we know the model itself is better than P-SpikeSSM, etc
  2. Also present a complete comparison table of energy, wall clock, and memory w.r.t. existing STOA models. I am unsure how much to trust only these neural signal decoding datasets for novelty.
Reviewer_biPcOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

When you record from the brain, different signals run on different clocks. For example, spikes are fast and bursty, local field potentials are slower, and dopamine slower still (over hundreds of milliseconds to seconds). A standard spiking neuron has a single forgetting rate, so it treats all of these signals the same way, and does not account for the fact that useful structure is spread across different timescales.

The proposed fix, MRD-SNN, gives each neuron several internal memory branches instead of one. Each branch updates on its own clock and remembers over its own window, so one branch can follow fast detail while another holds a slow trend. Each branch also keeps its history in a compact form, as a Legendre Memory Unit, which captures the shape of a recent window with a few numbers rather than storing the raw samples. When the neuron fires, it resets the part that spikes but leaves the branch memories untouched, so the slow states that remember outlive the fast event that fires. A dedicated input layer handles the fact that the modalities arrive on different clocks, counting only the channels that actually reported a fresh reading at each moment rather than filling in zeros or holding stale values.

The authors report gains on three cross-modal decoding tasks. On the most timescale-diverse dataset, a spike, LFP and dopamine reward task from a primate recording, MRD-SNN reaches about 78% against about 69% for the best baseline, with smaller margins of roughly 4 points on two public datasets. Ablations show that (a) a fast branch paired with a slow branch beats two fast or two slow branches at the same capacity (Table 4), and (b) fixing the memory budget and spreading it across many small branches beats concentrating it in one large memory (Table 6), and (c) giving each neuron its own private memory beats sharing memory across a layer (Table 5). Baselines are parameter-matched.

Critical strengths
  1. The paper shows that spreading a fixed memory budget across several branches helps: Table 6 holds the total memory budget fixed and shows accuracy improving as the number of branches grows, by about ten points on the hardest task. Table 4 points the same way, with mixed fast-and-slow pairs beating matched same-rate pairs on SHD and sMNIST. These are informative comparisons. They do not by themselves separate the benefit of different update rates from the benefit of different memory horizons, since holding the total budget fixed still lets the number of branch projections, the readouts, and the spread of horizons grow together, but the headline effect that many branches beat one is clear and well supported.

  2. The application is a good fit for the idea: Multimodal neural recordings really do arrive on different physical clocks, which is a more convincing motivation for multi-rate, multi-horizon memory than the usual image or audio benchmarks. The input layer is thoughtfully built for this, designed to handle ragged, asynchronous sampling by counting only the channels that reported a fresh reading rather than imputing zeros or holding stale values, and keeping each channel's own normalization so that modality identity survives on the way in.

  3. The central mechanism is isolated and the ablations are easy to read: The per-neuron memory result (Table 5) demonstrates the empirical value of giving each neuron its own private memory rather than sharing it across a layer, which is a more careful claim than saying it proves increased expressivity. The design choice that a spike resets the firing part but not the memories is a crisp statement of what the architecture is for. These are the kind of controlled comparisons a reviewer can trust.

Critical weaknesses
  1. The most relevant prior work is neither cited nor compared against. The ELM Neuron (Spieler et al., ICLR 2024, not cited in the submission) already develops the idea that a single neuron can gain substantial temporal expressivity from multiple internal memory states with heterogeneous timescales and nonlinear synaptic integration, and its Branch-ELM variant groups synaptic inputs into branch-like structures. This overlaps substantially with the motivation for MRD-SNN, in particular its emphasis on private, heterogeneous memory within an individual neuron. MRD-SNN does introduce distinct mechanisms, including the Legendre memory, the explicit multi-rate refresh schedules, and the spiking dynamics, so the paper should clarify what these add beyond ELM and ideally run it as a direct baseline on these tasks.

  2. The experiments support multi-branch, multi-horizon memory more than they support multi-rate updating, which is what the title emphasizes. Two of the paper's own controls point this way. When the physical memory horizon is held fixed and only the update rate changes, accuracy moves by at most about one point and stays inside the run-to-run noise (Table 3). When every branch is instead switched to update on each tick while the horizons are preserved, accuracy again barely moves, about 76.2 against 76.1 on the hardest task, so the multiple rates mainly buy a compute saving of roughly 1.8 times rather than accuracy (Table 9, Appendix E). The comparison that does favor mixed rates, Table 4, holds the window-length parameter fixed, so a slower-updating branch also reaches further back in time. That comparison changes the update rate and the memory horizon together and cannot separate the two. What consistently tracks accuracy is the number of branches and the spread of their horizons (Table 6). On the paper's own evidence the gain comes from carrying many branches at diverse horizons rather than from running them on different clocks, so the multi-rate framing claims more than the results show.

  3. The input handling is not fully pinned down, and the branch wiring is never isolated. The asynchronous input story has a loose end that the paper presents as a strength. Section 3.1 says an absent observation is never treated as a measured zero or a repeated measurement, yet Appendix D forward-fills dopamine after its first valid sample, and the modality ablation in Appendix B zeroes absent modalities. It is not stated whether a forward-filled dopamine value is passed to the network as a fresh observation or held out by the observation mask. Since not imputing stale or zero values is sold as a design advantage, this should be reconciled. Beyond that, the input layer itself is never ablated, so the contribution of the asynchronous handling is not separated from the rest of the model. The branches are wired to fixed, contiguous slices of the input rather than to particular modalities, so nothing guarantees that a branch's clock matches the timescale of what it is fed, and the authors note that a branch can mix modalities within its receptive field. Any alignment between branch rate and modality has to be learned rather than built in, which sits awkwardly with the one-branch-per-timescale picture in the figures.

Questions for authors
  1. Will you position MRD-SNN against the ELM Neuron (Spieler et al., ICLR 2024), which shares your central idea of per-neuron, multi-timescale, branch-structured memory, and can you run it as a baseline on your decoding tasks? If you see your contribution as distinct, what is the sharpest statement of what multi-rate updating and the spiking packaging add over it?

  2. Can you compare a single large memory against multiple branches under full-rate updating, with the total memory budget matched and the physical horizons tuned independently for each configuration? Appendix G already shows a single memory losing on a synthetic task, but not under this controlled setting on the decoding benchmarks, and this comparison would isolate the benefit of spreading memory across several horizons from the benefit of different update schedules.

  3. The branch masks are fixed and modality-blind. Can you add an ablation that compares them against modality-aligned wiring, where each branch sees a single modality, to test whether matching a branch's clock to the timescale of what it is fed actually helps? It would also help to ablate the input layer against zero-filling or resampling to a common clock, so its contribution can be separated from the branch machinery. It would also help to state plainly how forward-filled dopamine is handled, since Section 3.1 and Appendix D read as different rules.

3 comments

Q1:

We thank the reviewer for highlighting this important related work. Using the authors’ public implementation, we added ELM and Branch-ELM comparisons on all three datasets, with identical inputs and splits, approximately matched parameter counts, and five folds with three seeds per fold.

ModelSpike–LFP–DA (%)DANDI 000121 (%)DANDI 000447 (%)
ELM74.39 ± 6.9286.93 ± 1.6780.13 ± 3.22
Branch-ELM75.44 ± 6.2083.23 ± 1.3972.88 ± 3.47
MRD-SNN77.67 ± 4.8987.34 ± 1.0977.91 ± 3.12

MRD-SNN achieves higher mean test accuracy than ELM on Spike–LFP–DA and DANDI 000121 by 3.28 and 0.41 percentage points, respectively. On DANDI 000447, ELM leads by 2.22 points. MRD-SNN outperforms Branch-ELM on all three tasks by 2.23, 4.11, and 5.03 percentage points, respectively. These results show competitive performance, although MRD-SNN does not outperform ELM on every task.

MRD-SNN differs in three specific ways:

  1. Structured historical representation. Each dendritic branch summarizes history in a Legendre state, with an explicitly configurable nominal physical horizon Hk=nkθkΔtH_k=n_k\theta_k\Delta t that relates its memory span to the actual timescale of the neural signal.

  2. Independent control of memory horizon and refresh frequency. When nkn_k changes, θk\theta_k can be adjusted to preserve the horizon, while cached readouts remain available between refreshes. In the new matched-horizon controls, mixed refresh is only about 0.7 and 0.4 percentage points below high-rate refresh on SHD and sMNIST, respectively, while reducing nominal branch-update counts by approximately 45%. This demonstrates the value of multirate scheduling for controlling computational cost.

  3. Separation of spike generation and history retention. The soma integrates the direct-path current and dendritic memories to generate spikes. A spike resets only the soma, leaving dendritic histories intact so that long-term context persists across successive spikes.

Q2:

We appreciate the reviewer’s suggestion to separate memory allocation from update scheduling. We added comparisons between a single long memory and eight memories with different horizons under every-step updating, together with a multi-rate configuration using the same eight horizons.

All conditions use the same five folds, three seeds per fold, 100 training epochs, and validation-selected checkpoints. Each layer has Kq=80Kq=80, giving 2×128×80=20,4802\times128\times80=20{,}480 memory-state variables. The first two configurations update every step; the third preserves the eight physical horizons and uses intervals [2, 5, 8, 10, 12, 15, 20, 30]. All horizons are fixed.

ConfigurationKqSpike–LFP–DA (%)DANDI 000121 (%)DANDI 000447 (%)
Single long memory, full rate18073.48 ± 2.3683.78 ± 1.6173.85 ± 2.76
Eight distinct horizons, full rate81077.59 ± 4.4287.08 ± 1.4878.05 ± 1.62
Eight distinct horizons, mixed rates81076.99 ± 4.7886.02 ± 1.0578.55 ± 3.02

The eight full-rate branches improve over one long memory by 4.11, 3.30 and 4.20 points on the three tasks.

The matched-horizon mixed-rate model is within 0.60, 1.06 and 0.50 points of the full-rate eight-branch model, while reducing nominal branch updates by 85.54%, 85.91% and 85.63%, respectively. Thus rate diversity is primarily a scheduling and compute choice in this control; it is not required to explain the allocation gain.

For a fixed horizon, H=nθΔtH=n\theta\Delta t, a branch refreshed every tick requires a larger θ\theta. At fixed memory order and sufficiently large θ\theta, the discretized input matrix satisfies Bθ≈B0/θB_\theta\approx B_0/\theta, so each update writes less new information. This can make recent details harder to preserve. Appendix G illustrates performance failures when one high-rate branch must retain very long history while distinguishing recent timing. We therefore assign different memory horizons to different branches to reduce the burden on any single branch, and refresh long-memory branches less frequently to reduce computation.

Q3:

Thank you for these suggestions. We added controls for modality-aligned wiring and input handling, and clarify the original dopamine preprocessing below.

In the modality-aligned experiment on Spike–LFP–DA, each first-layer dendritic branch receives only its assigned modality: n = 2, 5 for Spike; n = 8, 10, 12, 15 for LFP; and n = 20, 30 for DA. The direct pathway still integrates all modalities, and the second layer receives fused hidden activity. With the same number of first-layer connections and the same total parameter count, modality-aligned wiring achieves 76.95 ± 5.58%, compared with 76.55 ± 4.92% for the original contiguous local wiring. The 0.40-percentage-point difference does not show a clear advantage from prescribing a modality–clock correspondence.

To distinguish the contributions of input handling and model architecture, we compared three input conditions for MRD-SNN, DH-LIF, ordinary LIF (FSNN), and ALIF. All conditions use the same five folds, three seeds per fold, 100 training epochs, and validation-selected checkpoints, with matched initialization within each model and seed. Accuracy is reported as mean ± SD (%).

Input protocolMRD-SNNDH-LIFFSNNALIF
Strict asynchronous77.35 ± 5.5871.93 ± 8.7666.55 ± 5.8465.39 ± 7.08
Raw zero fill66.06 ± 5.0750.96 ± 1.7856.27 ± 5.1550.47 ± 2.31
Uniform interpolation73.45 ± 5.3970.49 ± 7.5765.48 ± 5.3166.79 ± 4.95

MRD has the highest mean accuracy under all three input conditions. Strict asynchronous input improves MRD by 11.29 points over raw zero filling and 3.90 points over uniform interpolation. With the same asynchronous input, MRD remains 5.42 points above DH-LIF and 10.80 above FSNN. Thus input handling contributes to performance but does not account for the entire gap between models. The advantage that persists under identical input conditions supports an important contribution from dendritic memory branches.

Regarding DA forward filling, DA is sampled every 100 ms. Between sampling times, the network does not repeatedly receive the previous DA value. However, if a DA value is missing at a scheduled sampling time, the original preprocessing fills it with the most recent valid value and supplies it to the network at that time. Appendix D therefore describes the treatment of missing samples, whereas Section 3.1 omitted this exception. We will reconcile the two descriptions.

Dear Reviewer,

thank you again for your thoughtful comments. We have posted our responses and additional experimental results above. As the discussion period is coming to an end, we would appreciate it if you could take a look when convenient and let us know if any points remain unclear. Thank you for your time.

Comments on the paper

1 comment

Dear Reviewer,

thank you very much for carefully considering our responses and for updating your score. We sincerely appreciate your time and constructive feedback, which helped us present the contributions of our work more clearly.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.