{"id":"c4c07d59-7123-4b49-9a75-b9ea37adf3ed","arxiv_id":"2506.01117","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":2,"one_line_summary":"STDL trains spiking neural networks by partitioning them into locally supervised subnetworks with capacity-optimized auxiliary networks and online temporal updates, matching BPTT accuracy with substantially lower GPU memory.","lead":"The paper introduces STDL, a training method for spiking neural networks that splits the network into independently trained chunks with auxiliary networks and drops most temporal gradient dependencies. It reports accuracy close to full backpropagation-through-time on seven vision datasets while using roughly four times less GPU memory on ImageNet.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"ImageNet GPU-memory comparison is not at a fixed batch size: BPTT used batch 256 (Appendix B) while STDL used 512, so the headline 4.0x saving is an uncontrolled ratio and likely understates the true factor.","rationale":"Among the issues the reader raised (informal Proposition 2, missing error bars, batch-size mismatch), the batch-size mismatch is the most load-bearing because it directly concerns the quantitative half of the central claim: 'significant GPU memory savings by 4.0x-4.7x.' The gradient-conflict issue in the auxiliary networks is a legitimate concern about the method's mechanism, but it does not threaten the empirical validity of the reported results: the method achieves the stated accuracy on seven datasets regardless of the theoretical explanation. The batch-size inconsistency, by contrast, means the headline memory ratio is not a controlled measurement; it could be off by a factor of roughly two. This is an internal inconsistency (Table II header vs Appendix B) rather than a disagreement with consensus, and it is directly checkable. The reader identified this issue in the rationale but chose a different weakest assumption; I agree with the conditional verdict, and the proposed test would let the authors correct the reported factor.","tokens_in":25893,"tokens_out":10797,"duration_ms":104762,"concrete_test":"Re-measure peak GPU memory for BPTT and STDL on ImageNet (SEWResNet-34, T=4) at the same batch size. If BPTT does not fit at batch 512, use gradient accumulation with a smaller physical batch to emulate a 512-observation update and record the peak memory for one training step; then compare with STDL at the same effective batch size. If the ratio comes out near 4.0x, the reported comparison is accidentally fair; if it is near 8x, the headline 4.0x is an artifact of the batch-size mismatch and the text and Table II must be corrected.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Table II reports a 4.0x GPU-memory reduction for STDL vs BPTT on ImageNet under a caption stating that memory is measured at batch size 512 for all methods. Appendix B, however, states that BPTT on ImageNet was run at batch size 256 because of memory constraints, while STDL used batch size 512 (Table VII). The reported BPTT memory of 424.35 GB is therefore at batch 256, whereas the STDL memory of 106.53 GB is at batch 512. Because activation memory scales roughly linearly with batch size, the central memory-savings claim is computed from mismatched operating points. A like-for-like comparison would put BPTT at approximately 848 GB at batch 512 (or STDL at about 53 GB at batch 256), i.e. roughly 8x savings rather than 4.0x. The qualitative claim that STDL saves substantial memory is not in doubt, but the specific 4.0x-4.7x factor in the abstract and Section V-A2 is not supported by the reported protocol, and the paper is internally inconsistent between Table II and Appendix B.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes spatio-temporal decoupled learning (STDL), a training method for spiking neural networks that partitions a network into subnetworks, trains each subnetwork with an auxiliary network assembled from subsequent layers under a memory budget, and omits temporally dependent gradients to enable online updates. The authors provide a greedy partitioning algorithm with an optimality argument, an information-theoretic motivation for auxiliary network construction, and experiments on seven datasets (CIFAR-10/100, SVHN, ImageNet, CIFAR10-DVS, GAIT-DAY-DVS, HAR-DVS). The headline claims are that STDL matches BPTT accuracy while reducing GPU memory by 4.0x-4.7x, and that it is the first local-learning-style SNN training method to scale to ImageNet at BPTT-level accuracy.","tokens_in":26131,"tokens_out":6624,"duration_ms":69250,"significance":"If the claims hold, this is a significant advance for memory-efficient SNN training: the method combines a local-learning memory profile with accuracy close to BPTT across diverse architectures and datasets, and it ships code and detailed training configurations. The accuracy numbers are internally consistent across Tables II and III, and the ablations in Section V-E support the importance of auxiliary capacity. However, the headline memory ratio rests on an uncontrolled batch-size comparison, the central theoretical proposition is not actually proved, and the ImageNet comparison lacks error bars; these issues must be resolved before the stated contributions can be accepted.","major_comments":[{"comment":"The image-classification table states in its caption that all GPU memory is measured at batch size 512, but Appendix B reports that BPTT on ImageNet was run at batch size 256 because of memory constraints, and Table VII also distinguishes the ImageNet finetune batch size from the baseline batch size. The reported BPTT memory of 424.35 GB therefore corresponds to batch 256, while the STDL memory of 106.53 GB corresponds to batch 512. Since activation memory scales roughly linearly with batch size, the 4.0x ratio cited in the abstract and Section V-A2 is not computed at a fixed operating point; at a common batch size the true ratio would be approximately 8x (at batch 512) or 2x (at batch 256). The authors should re-measure at a fixed batch size or clearly report the corrected factor, because the exact memory saving is a central claim of the paper.","section":"Table II, Appendix B, and Table VII"},{"comment":"Proposition 2 is not a valid proof. The construction of a function-preserving expanded network via zero-padding and identity residual connections is not justified for the actual auxiliary networks, which include spiking activations, downsampling layers, and normalization; moreover, the step claiming that an expanded network has 'the potential to discover' a more informative representation is an assertion rather than a derivation. Since this proposition is used to motivate the depth/width maximization, the authors should either replace it with a theorem under explicit assumptions or clearly present the capacity-growing rule as a heuristic that is empirically supported by the ablations in Section V-E.","section":"Section IV-C, Proposition 2"},{"comment":"The auxiliary network for each subnetwork is built by reusing actual subsequent layers (including the classifier), and those same layers are later trained as part of later subnetworks. The stop-gradient operator in Eq. (10) blocks gradient flow from later subnetworks into the current subnetwork's representation, but it does not prevent the local loss of subnetwork k from updating the parameters of its auxiliary network G^k, which are shared with later subnetworks. The paper does not analyze or ablate this overlapping-gradient coupling. An explicit experiment, such as training the auxiliary network with a detached copy or frozen weights, or measuring gradient conflict between the overlapping losses, is needed to support the claim that the reuse is beneficial rather than harmful.","section":"Sections IV-A, IV-C, and V-E"},{"comment":"The central claim of BPTT-comparable accuracy on ImageNet rests on single runs (BPTT 70.12 vs STDL 69.87, with no standard deviation), and the HAR-DVS rows in Table IV also lack error bars. Given that the other datasets show seed-to-seed variation on the order of 0.1-0.2 percentage points, a single ImageNet run is insufficient evidence to support the phrase 'comparable accuracy to BPTT' on large-scale data. The authors should report multiple seeds, or at minimum provide a clear statistical justification for the single-run comparison.","section":"Section V-A, Table II; Section V-B, Table IV"}],"minor_comments":[{"comment":"The y-axis labels and legend in Figure 9 are rendered as unreadable encoded text in the manuscript; please ensure the figure is regenerated with correct fonts.","section":"Figure 9"},{"comment":"The statement that temporally dependent gradients 'contribute negligibly' is supported only by an exponential-decay argument and a citation to prior work; reporting an empirical measurement of the omitted-gradient norm would make the approximation more convincing.","section":"Section IV-D"},{"comment":"The comparison with published BPTT-based methods uses results from their original papers under different training settings (e.g., different numbers of time steps); a sentence noting that this is an informal comparison would help readers interpret the table.","section":"Table III"},{"comment":"The symbol rho is defined as a memory efficiency ratio but is never used elsewhere in the text; either explain its role or remove it to avoid confusion.","section":"Table VIII"}],"recommendation":"major_revision","confidential_remarks":"The manuscript is within the scope of the journal, and the empirical accuracy results appear internally consistent. The main concern for the editor is that the headline 4.0x memory saving is likely an understatement once batch sizes are aligned, but the current reporting is internally inconsistent and must be corrected; this is a substantive but fixable issue. I do not see circularity in the evaluation, since the method is compared against external BPTT baselines. The theoretical section needs significant tightening before the paper can be accepted."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: STDL is a genuine step forward for local/online SNN training. The combination of greedy subnetwork partitioning under a memory budget, auxiliary networks built by reusing subsequent layers to maximize capacity, and temporal gradient omission is new as a package, and the experiments are unusually thorough: seven datasets, multiple architectures, neuron models, ablations, representation analysis, and code. The accuracy numbers are consistent with the claim that STDL matches BPTT while cutting memory. I buy the central empirical result.\n\nWhat's actually new: the greedy partitioning criterion (minimize subnetwork count under a memory constraint) and the capacity-maximizing auxiliary network construction are not in the cited literature. The ablation studies show both components matter: replacing the auxiliary structure with generic convs or uniform layer-wise partitioning hurts accuracy. The representation alignment analysis (CKA, linear separability) supports the weak coupling story. This is a solid engineering contribution.\n\nSoft spots, in order:\n\n1. The ImageNet memory comparison is not like-for-like. Table II's caption says memory is measured at batch size 512, but Appendix B says BPTT on ImageNet ran at batch 256 due to memory limits, while STDL used 512. So the reported 4.0x saving (424.35 GB vs 106.53 GB) compares BPTT at 256 with STDL at 512. The stress-test note is right: like-for-like, the factor is roughly 8x, not 4x. The qualitative claim of substantial savings is fine, but the abstract and Section V-A2 repeat 4.0x-4.7x as if it were controlled. This needs to be corrected or at least clearly footnoted.\n\n2. Proposition 2 is not a rigorous proof. The argument that an expanded network 'has the potential to discover a more informative representation' doesn't establish monotonicity of mutual information after training. It's a plausibility argument, not a theorem. The DPI-based surrogate in Section IV-C is also loose: maximizing a lower bound on I(s_local;y) doesn't directly maximize I(s_local;y). Neither flaw breaks the method, but the paper overstates the theoretical backing.\n\n3. Some headline results lack error bars: ImageNet (Table II) and HAR-DVS (Table IV) report single runs. Minor, since the differences to BPTT are small and consistent elsewhere.\n\n4. The interaction between auxiliary networks and later subnetworks is unexamined. Since auxiliary networks reuse subsequent layers that are later trained as part of later subnetworks, those shared parameters receive multiple local losses. The paper doesn't analyze or ablate this. It apparently works, but it's a real gap.\n\nOverall: this is a worthwhile paper that deserves serious refereeing. The fixes are addressable: correct the memory comparison, tone down the proof claims, add error bars. I hope it gets revised rather than rejected.","headline":"Solid empirical paper on memory-efficient SNN training, with a real but fixable inconsistency in the headline ImageNet memory number.","tokens_in":26656,"tokens_out":2931,"would_cite":true,"duration_ms":28323,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["68T07"],"pacs":[],"model":"deepseek-v4-flash","headline":"Spatio-temporal decoupled learning trains spiking neural networks at BPTT accuracy with 4x lower GPU memory by splitting the network into subnetworks guided by auxiliary networks built from their own subsequent layers.","keywords":["spiking neural networks","local learning","backpropagation through time","memory-efficient training","online learning","auxiliary networks","network partitioning","surrogate gradients"],"falsifier":"Train STDL but freeze the auxiliary network parameters, or detach the shared downstream layers so earlier subnetworks cannot update them, while keeping everything else identical. If accuracy stays at BPTT level, the overlap is not the source of coupling; if accuracy drops toward ELL's level, the shared-parameter updates are load-bearing.","tokens_in":25671,"feed_emoji":"⚡","tokens_out":4326,"duration_ms":44277,"temperature":0.7,"pith_summary":"The paper proposes STDL, a training method for spiking neural networks that splits the network into independently trained subnetworks while still matching the accuracy of full backpropagation through time (BPTT). The payoff is memory: on ImageNet, training SEWResNet-34 with STDL uses roughly 4x less GPU memory than BPTT at comparable accuracy, and the savings grow with the number of time steps. The method works because spatial decoupling is done carefully, using a minimal number of subnetworks under a memory budget and guiding each one with an auxiliary network built from actual subsequent layers, while temporal decoupling drops only gradients whose contribution decays exponentially. The paper argues this is the first local-learning-style SNN training that scales to ImageNet without an accuracy penalty.","feed_headline":"Spiking nets hit BPTT accuracy at 4x less memory","feed_subtitle":"STDL splits the network into subnetworks with auxiliary guides, matching full backprop while freeing most cached states.","key_machinery":"The load-bearing mechanism is the pairing of each subnetwork with an auxiliary network constructed by selecting the largest subset of its subsequent layers under the memory constraint, combined with a greedy partition that minimizes the number of subnetworks. The auxiliary network provides a learned, structure-shaped target so that earlier layers' local losses align their representations with BPTT; the greedy partition keeps subnetworks large enough to preserve synergy. Temporal decoupling then computes gradients online via Eq. (18), using only quantities available at time t and discarding temporally dependent terms whose influence decays like $\\lambda^{t'-t}$.","core_discovery":"STDL establishes that decoupling both space and time in SNN training need not cost accuracy. Spatial decoupling partitions the network into K subnetworks trained by local losses; a greedy partition minimizes K under a memory constraint, and each subnetwork is paired with an auxiliary network that reuses a largest-capacity subset of its subsequent layers, steering its representations toward those BPTT would produce. Temporal decoupling discards temporally dependent gradients, which decay as $\\lambda^{t'-t}$, enabling online updates and making memory independent of the number of time steps. Across seven datasets and several architectures, STDL matches or slightly exceeds BPTT accuracy while cutting GPU memory by 4.0x-4.7x on static image tasks and by up to 20.8x on long-time-step event-based tasks.","pith_inferences":["A natural untested ablation would freeze the auxiliary network parameters or detach the shared downstream layers so earlier subnetworks cannot update them; the paper does not isolate whether the coupling benefit comes from representational alignment or from direct multi-loss optimization of overlapping parameters.","Since the temporal-gradient cutoff relies on exponential decay with factor $\\lambda$, networks with slower membrane decay may need to retain some temporal terms; a threshold based on $\\lambda$ rather than omitting all such terms is a testable extension.","The partition-plus-auxiliary recipe is not obviously specific to spiking neurons and could be applied to other stateful sequence models where BPTT memory is the bottleneck, provided a decay-like structure justifies dropping long-range temporal gradients.","An adaptive memory constraint, loosened as training stabilizes, could further close the small accuracy gap on ImageNet while preserving most of the memory savings."],"forward_implications":["STDL-trained spiking networks can fit on GPUs that cannot hold BPTT's cached states, enabling larger architectures or longer temporal windows at the same memory cost.","Because memory use is independent of the number of time steps, STDL is especially favorable for event-based data with high temporal resolution, where BPTT memory grows linearly with time steps.","The method transfers across spiking neuron models: on PLIF and ALIF neurons it achieves 14x-24x memory savings while maintaining BPTT-level accuracy.","Representation analysis with CKA and linear probing shows STDL layers align closely with BPTT's, suggesting the accuracy parity reflects genuinely similar learned features rather than a lucky loss landscape.","The auxiliary networks are used only during training and discarded at inference, so the deployed model is a standard spiking network with no extra cost."],"supporting_citations":[{"why":"Defines the BPTT surrogate-gradient training baseline and the gradient formulas whose memory cost STDL aims to reduce.","marker":"[14]"},{"why":"DECOLLE is the fixed-random-auxiliary supervised local learning baseline that STDL outperforms by large margins.","marker":"[27]"},{"why":"ELL provides the trainable linear-classifier local learning baseline and the uniform-partition setup used to expose the weak coupling issue.","marker":"[18]"},{"why":"SLTT is the temporal-only decoupling baseline, supporting the claim that temporally dependent gradients contribute negligibly.","marker":"[59]"},{"why":"Supplies the SEWResNet-34 architecture used for the ImageNet scaling experiments.","marker":"[28]"},{"why":"Provides the information-theoretic data processing inequality used to justify maximizing auxiliary network capacity.","marker":"[66]"},{"why":"CKA is the representation-similarity measure used to show STDL layers align with BPTT's.","marker":"[71]"}],"fun_headline_variants":["STDL decouples SNN training, matches BPTT at 4x less memory","Spatio-temporal decoupling: SNN accuracy without BPTT cost","SNN training: decouple to beat memory wall, keep accuracy","STDL: split SNNs, cut GPU memory 4x, keep BPTT accuracy","Decoupling SNN training: efficient online, BPTT-grade results"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The whole accuracy claim rests on the assumption that when an early subnetwork's local loss updates a layer that a later subnetwork also trains, those overlapping updates work together instead of pulling the layer in different directions.","fun_headline_variants_meta":{"raw":{"variants":["STDL decouples SNN training, matches BPTT at 4x less memory","Spatio-temporal decoupling: SNN accuracy without BPTT cost","SNN training: decouple to beat memory wall, keep accuracy","STDL: split SNNs, cut GPU memory 4x, keep BPTT accuracy","Decoupling SNN training: efficient online, BPTT-grade results"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000181,"raw_usage":{"total_tokens":1322,"prompt_tokens":978,"completion_tokens":344,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":594,"completion_tokens_details":{"reasoning_tokens":238}},"tokens_in":594,"tokens_out":344,"duration_ms":4144,"temperature":1.0,"reasoning_tokens":238,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T11:50:53.873293+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Train STDL but freeze the auxiliary network parameters, or detach the shared downstream layers so earlier subnetworks cannot update them, while keeping everything else identical. If accuracy stays at BPTT level, the overlap is not the source of coupling; if accuracy drops toward ELL's level, the shared-parameter updates are load-bearing.","supporting_citations":[{"cited_title":"Deep spike learning with local classifiers,","cited_arxiv_id":null,"evidence_quote":"ELL provides the trainable linear-classifier local learning baseline and the uniform-partition setup used to expose the weak coupling issue."},{"cited_title":"Towards memory-and time-efficient backpropagation for training spiking neural networks,","cited_arxiv_id":null,"evidence_quote":"SLTT is the temporal-only decoupling baseline, supporting the claim that temporally dependent gradients contribute negligibly."},{"cited_title":"Similarity of neural network representations revisited,","cited_arxiv_id":null,"evidence_quote":"CKA is the representation-similarity measure used to show STDL layers align with BPTT's."}],"review_version":1}