- Rating
- 2
- Confidence
- 2
- Informativeness
- 2
- Presentation
- 3
- Soundness
- 1
Read review
Paper summary
The authors propose DPET, a method for training recurrent SNNs with eligibility traces as opposed to BPTT. The method reduces the memory and computation required for storing eligibility traces through one additional approximation: replacing the dynamics Jacobian with the average of its trace (i.e., a scalar). They train their system on image benchmarks and measure its memory usage and compute time.
Critical strengths
- Lowering the resources required to train an SNN relative to BPTT, or even e-prop, is a sensible goal.
- The proposed modification to e-prop both lowers the memory and compute.
- They have an extensive set of empirical evaluations on image benchmarks.
Critical weaknesses
- If the primary goal is memory reduction for training an SNN on a GPU, then activation checkpointing should be the BPTT baseline for memory usage. This has the added benefit of not needing to redesign batchnorm and achieves the exact same accuracy as full BPTT.
- There are no tasks reported that actually require temporal dynamics, such as SHD or SSC, where the method will actually be exercised relative to BPTT. All evaluations are on static images with 4-10 time steps, involving nearly no need for temporal credit assignment.
- e-prop is already an approximation to BPTT, so adding in another approximation would seem to require some additional error analysis, which seems like it belongs in the main text.
Questions for authors
- If lagged temporal distillation doesn't help with missing future pathways, how does it help?
- The bold rows in table 7 are misleading. The fair comparison is torch vs. torch and should be portrayed that way.
- Figure 2's legend is a little confusing.
- Do you introduce the membrane potential regularization? Or is that an existing method with some modifications?
8 comments
Response to Reviewer TQpe (cont.)
W2. There are no tasks reported that actually require temporal dynamics, such as SHD or SSC, where the method will actually be exercised relative to BPTT. All evaluations are on static images with 4-10 time steps, involving nearly no need for temporal credit assignment.
The 4–10 step range applies only to our main static-image benchmarks; the original manuscript already included Sequential CIFAR-100 at (presenting one image column per step; Section 5.3, Table 6), where DPET (%) and DPET+LTD (%) clearly outperform SLTT (%) and approach BPTT+TET (%).
Even so, we would like to clarify the role of temporal benchmarks in our study. Online learning rules must approximate historical sensitivities, omit future feedback paths, and cannot use cross-time-step operations (e.g., temporal convolution), which inherently limits raw sequential modeling capacity compared with full BPTT. Hence, online learning is not designed to replace BPTT on tasks with rich temporal dependencies. Instead, temporal tasks serve as stress tests for temporal credit assignment, because optimization on these tasks is highly sensitive to temporal gradients, and clear gains over other online rules directly validate the fidelity of DPET's eligibility trace.
Following your suggestion, we also evaluated DPET on SHD under two complementary settings (also reported in our response to Reviewer VEDm, W1). All results are test accuracy, mean ± SD over three seeds, selected by best validation accuracy, without LTD or membrane regularization.
1. Feedforward LIF on SHD (): isolating within-layer temporal credit assignment. To directly test whether the eligibility trace preserves historical sensitivity across long horizons, we first use a two-layer feedforward LIF network (128 units per layer, fixed decay 0.95, , 120 epochs), where no recurrent weights exist and temporal credit must propagate entirely through membrane dynamics.
| Rule | BPTT | SLTT | OTTT | NDOT | DPET |
|---|---|---|---|---|---|
| SHD test acc. (%) | 81.38 ± 1.30 | 61.41 ± 0.11 | 76.30 ± 1.08 | 70.77 ± 0.55 | 79.78 ± 1.05 |
Without historical eligibility (), SLTT drops to %. DPET achieves the highest accuracy among all online rules (%, outperforming OTTT and NDOT) and closes most of the gap to BPTT (%), demonstrating that projecting postsynaptic dynamics onto the eligibility trace effectively preserves long-horizon temporal credit.
2. Recurrent LIF on SHD (): evaluating performance in a higher-capacity recurrent model. We next evaluate a two-layer recurrent LIF architecture (1024 units per layer, learned decays, , 100 epochs, per-step cross-entropy), extending DPET with eligibility traces for recurrent weights.
| Rule | SHD test acc. (%) |
|---|---|
| BPTT | 87.13 ± 0.79 |
| SLTT | 84.41 ± 0.90 |
| OTTT | 85.96 ± 1.07 |
| NDOT | 85.03 ± 1.32 |
| pp-prop () | 86.81 ± 0.71 |
| Weighted DPET | 85.48 ± 1.21 |
| DPET | 86.75 ± 0.85 |
With recurrent synaptic connections, all rules achieve stronger accuracy. DPET reaches %, matching the performance of pp-prop . Together, the two SHD experiments show that DPET preserves long-range membrane-level credit.
Response to Reviewer TQpe (cont. 2)
W3. e-prop is already an approximation to BPTT, so adding in another approximation would seem to require some additional error analysis, which seems like it belongs in the main text.
Thank you for this suggestion. Your description is accurate: DPET adds one approximation to e-prop. Like e-prop, it omits feedback through future deeper states; in addition, it replaces the per-neuron eligibility traces with a single shared trace.
Consider one sample in a feedforward dense LIF layer with current input . For each output neuron , e-prop keeps an exact eligibility (the sensitivity of that neuron's membrane potential to its incoming weights, accumulated over past steps). It evolves with the neuron's own one-step Jacobian , as . DPET keeps one input-sized trace for the whole layer, with replaced by the layer mean , as . Their gap follows
The first term carries the earlier gap forward; the second is the new gap introduced at step . Since is the same vector for every , the sum of the squared norms of the newly introduced gaps is , which the mean minimizes. The newly introduced gap is zero when all neurons share the same Jacobian at that step.
With the same current spatial error signal , the gradient rows computed from DPET's trace and from exact eligibility differ by . Minimizing the new gap at each step does not guarantee the smallest accumulated difference (Appendix A.6), so we also measure it. Table 5 uses e-prop's exact eligibility as its exact local reference. Against it, DPET's error is 0.154, compared with 0.287–0.353 for SLTT, OTTT and NDOT; only pp-prop, which stores an additional postsynaptic trace, is lower. Against full BPTT+SG, where the error also includes the effect of omitted future feedback, DPET has the lowest error of these compact rules (0.342).
In the revision, we will move this analysis into Section 4.1, next to the projection objective.
Q1. If lagged temporal distillation doesn't help with missing future pathways, how does it help?
Thank you for the question. Without future feedback, the update at step is driven by the current loss ; there is no explicit target from later-step predictions. LTD supplies this on the target side: the step- output is trained to match the network's later-step predictions for the same class, collected from previously completed minibatches. It therefore needs no future Jacobians of the current sample, while gradients flow only through the current-step predictions.
Adding LTD to DPET raises mean accuracy from % to % on CIFAR-100 and from % to % on CIFAR10-DVS (Table 4). We will revise Section 4.2 to state this relation directly.
Response to Reviewer TQpe
Thank you for the suggestions on checkpointing and for asking us to clarify the approximation and auxiliary methods. Our responses are as follows.
We agree that checkpointed BPTT is an appropriate memory baseline. Gradient checkpointing (GC) recomputes activations during the backward pass instead of storing them, so it keeps the exact BPTT update and, as you note, needs no change to batch normalization. DPET, by contrast, requires BN after the spiking neuron or a projection without BN (Section 4.4).
For the BPTT baseline, we ran spatial GC in SpikingJelly's multi-step mode, where each layer processes all time steps before going on to the next layer. DPET is an online rule and processes one time step at a time through the whole network. We therefore report BPTT in both modes. All runs use SpikingJelly 2.0.0 on one RTX 5090, Spiking VGG11 for CIFAR-100 at T=6, batch size 128 and BF16. The BPTT+GC condition uses the level-4 automatic layout search with binary spike compression [1]. All conditions use the torch backend. Each condition runs with 10 warm-up and 100 timed iterations. We report the maximum peak allocation and the median time per batch.
At T=6, BPTT in multi-step execution mode is faster than BPTT in single-step mode, since the computation on stateless layers can be parallelized over the time axis. GC lowers the memory cost of BPTT while keeping the exact update at the cost of increasing the time per batch from 51.5 to 67.3 ms. DPET reduces peak memory by about 41.0% relative to GC with an update time close to single-step BPTT, at the cost of approximating the temporal paths that GC preserves. We will add this comparison to the paper.
Furthermore, spatial GC and DPET can be used together. Rather than checkpointing the multi-step BPTT graph, we checkpoint only a subset of spatial activations within each DPET time step and recompute them for that step’s local backward pass before advancing to the next step. This leaves DPET’s online update unchanged. Compared with pure DPET, DPET+GC (with spike compression) reduces peak memory by 16.8% and increases time per batch by 10.1%.