{"id":"de2e857e-c5fe-4b23-b2ef-1a53464f7fa6","arxiv_id":"2506.14984","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"A delay-shifted STDP rule that co-learns synaptic weights and delays improves MNIST classification accuracy in the Diehl-Cook spiking network compared to standard STDP and DR-STDP.","lead":"This paper introduces DS-STDP, a learning rule that trains both the strength and the delay of connections in spiking neural networks. On MNIST digit classification, networks using DS-STDP outperform networks trained with standard STDP or with the prior DR-STDP delay rule.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"DS-STDP's reported superiority is not yet stable: DR-STDP was deliberately weakened and accuracies are selected on the test set, so a fair re-run could change the ordering.","rationale":"The paper's DS-STDP rule is novel and the theoretical derivation is plausible, but the empirical comparison is the load-bearing part of the central claim. The reader's conditional verdict already identifies the deliberate weakening of DR-STDP as the weakest assumption, and the same concern is supported by the explicit statement in Section 3.3.1. The additional use of best test-set accuracy as an oracle and the absence of repeated runs compound the problem, because DS-STDP's best epochs are systematically later than those of the baselines. The proposed concrete test uses the available code and directly checks whether the reported ordering is a property of the learning rule or of the evaluation protocol. Since the reader's verdict is already CONDITIONAL and this stress-test does not identify a reason to change it, the verdict is left unchanged rather than escalated to rejection; the paper may be correct, but the headline empirical claim needs a fair, repeated, and validation-based evaluation to be substantiated.","tokens_in":13793,"tokens_out":4738,"duration_ms":47621,"concrete_test":"Using the released repository, rerun the 400- and 900-neuron rate-coding experiments with: (1) DR-STDP hyperparameters set equal to DS-STDP's Table 4 values (weight A+ = 5e-4, A- = -5e-6, tau = 20 ms; delay A'_+ = 1.2e-4, A'_- = -1.2e-2, tau' = 20 ms); (2) early stopping chosen on the held-out validation set rather than the test set; (3) at least 10 independent Poisson encodings/seeds per method, reporting mean and standard deviation of test accuracy. If DS-STDP no longer beats DR-STDP, or beats it by less than one standard deviation, the 'consistently superior' claim is not supported.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The headline claim that DS-STDP 'consistently achieves superior performance' depends on an evaluation that is not yet controlled for two confounds. First, Table 4 deliberately halves every DR-STDP amplitude and time constant relative to DS-STDP/STDP, with the authors stating in Section 3.3.1 that 'hyperparameters for DR-STDP are configured to slow the training in order to compensate for the non-sparse updates.' This makes DR-STDP a non-representative baseline, so the comparison may reflect tuning rather than the rule. Second, Section 3.3.1 also states: 'We use the best testing accuracy for a given epoch to simulate early halting under ideal conditions.' Selecting the best accuracy on the test set is an oracle criterion, and because DS-STDP peaks at later epochs (Tables 6-7: epoch 78-99 for DS-STDP versus 10-56 for DR-STDP and STDP in most configurations), oracle selection can disproportionately inflate DS-STDP's advantage. With a single run per configuration and no error bars, margins of 0.006-0.02 accuracy may fall within run-to-run or Poisson-sampling noise. The central claim would only be secure if the advantage survives a fair DR-STDP configuration, validation-based early stopping, and repeated seeds.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper introduces delay-shifted spike-timing dependent plasticity (DS-STDP), an extension of STDP in which pre-synaptic spike trains are shifted by a learned delay and the delay update uses the STDP kernels with the potentiation and depression terms interchanged. The authors reformulate the rule in trace form, compare it analytically to DR-STDP, and evaluate it on the unsupervised Diehl-Cook MNIST benchmark with models of 100 to 900 excitatory neurons, reporting that DS-STDP consistently outperforms DR-STDP and STDP. They also introduce a latency-based scoring function called spike responsiveness and analyze how trainable delays change training dynamics and parameter efficiency.","tokens_in":14056,"tokens_out":3339,"duration_ms":35394,"significance":"If the empirical claim held, DS-STDP would be a useful local, biologically plausible rule for co-learning synaptic weights and delays, and the paper would offer a clear analytical bridge between STDP and delay learning. The derivation is internally coherent, the comparison to DR-STDP is instructive, and the authors provide a public code repository. However, the central empirical claim is not yet fully supported: the DR-STDP baseline is deliberately tuned to train more slowly, test-set accuracy is used as an oracle early-stopping criterion, and no repeated-seed statistics are reported. These issues are fixable and do not undermine the rule's derivation, but they currently prevent the paper from supporting its headline that DS-STDP consistently achieves superior performance.","major_comments":[{"comment":"The comparison against DR-STDP is confounded by deliberate asymmetric hyperparameter tuning. The text states that 'hyperparameters for DR-STDP are configured to slow the training in order to compensate for the non-sparse updates,' and Table 4 confirms that every DR-STDP amplitude and time constant is half the corresponding DS-STDP/STDP value. Because the headline claim is that DS-STDP consistently outperforms DR-STDP, the authors must either run DR-STDP with the same weight and delay update amplitudes/time constants as DS-STDP, or perform a hyperparameter sweep for DR-STDP under the same selection protocol, and show that the reported ordering is robust.","section":"Section 3.3.1, Table 4"},{"comment":"The use of 'the best testing accuracy for a given epoch to simulate early halting under ideal conditions' is an oracle criterion that can systematically favor DS-STDP. DS-STDP reaches its reported peak accuracy at later epochs (e.g., epoch 78 or 99 for the 625- and 900-neuron models) than DR-STDP or STDP, so selecting the best test-set accuracy over all epochs can inflate DS-STDP's advantage. The authors should instead select the model using the held-out validation set (which they already created) or report accuracies at a common fixed epoch, and then evaluate the selected model on the test set.","section":"Section 3.3.1, Tables 6-7"},{"comment":"No repeated runs or error bars are reported, although the input spike trains are Poisson-sampled and the authors note that nondeterministic CUDA operations can introduce variation. The reported accuracy margins between DS-STDP and DR-STDP/STDP are often 0.006-0.02, which may be within run-to-run or sampling noise. The central superiority claim requires multiple seeds with mean and standard deviation (or paired comparisons), especially because the experimental setup already acknowledges sources of nondeterminism.","section":"Section 3.3 and Section 4, Tables 6-8"}],"minor_comments":[{"comment":"The phrase 'against another existing methods' should read 'against other existing methods'.","section":"Abstract"},{"comment":"Equation (20) uses Iverson-bracket notation without defining it; the condition 's = floor(d(s)/Delta t)' should be explicitly introduced as an indicator function.","section":"Section 2.3, Equations 19-20"},{"comment":"The caption 'using the better performing method for each configuration' should clarify which scoring function (rate or responsiveness) is selected and why; as written, it is unclear whether the choice is made using test-set performance, which would compound the oracle-selection issue.","section":"Section 4, Figure 6"},{"comment":"The five-number summaries of per-class accuracy are useful, but the figure would benefit from stating the number of classes and the aggregation across runs; as presented, it is not possible to tell whether the box statistics are computed from a single run or multiple runs.","section":"Section 4, Figure 7"},{"comment":"The spike responsiveness score Qn is defined in Equation (21) but the text introduces it without naming the symbol; please state explicitly that Qn denotes spike responsiveness before or immediately after the equation.","section":"Section 3.2, Equations 21-23"}],"recommendation":"major_revision","confidential_remarks":"The paper is within the scope of a neuromorphic or computational neuroscience venue, and the code release is a strength. The main risk is not the derivation but the evaluation protocol: the DR-STDP baseline is intentionally weakened, the early-stopping criterion uses test-set labels, and the runs are single-shot. If the authors can add a fair DR-STDP configuration, validation-based early stopping, and repeated-seed statistics, the central claim could become publishable. If those additions are not feasible, the empirical claims should be substantially softened."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Start with the bottom line: DS-STDP is a genuine extension of STDP—shift the pre-synaptic trace by the learned delay and flip the kernels for the delay update. The analytical comparison to DR-STDP is the strongest part: the single-pair updates are identical in value but fire at different times, and the sparsity analysis explains a real behavioural difference. The rule is novel, coherent, and the code is available. Related-work coverage is adequate, and the reliance on [NG24] for the sign of delay updates is explicit.\n\nThe soft spot is the empirical claim of “consistent superiority.” Section 3.3.1 admits that DR-STDP hyperparameters are “configured to slow the training” to compensate for non-sparse updates—every amplitude and time constant is halved relative to DS-STDP/STDP. That is not a representative baseline; it is deliberately handicapping the comparison. The same section says they use the best testing accuracy over epochs to “simulate early halting under ideal conditions,” i.e., oracle selection on the test set. Because DS-STDP peaks much later (epochs 78–99 vs 10–56 for DR-STDP/STDP), oracle selection can disproportionately inflate its advantage. With one run per configuration and no error bars, the 0.006–0.02 accuracy margins sit inside plausible run-to-run and Poisson-sampling noise.\n\nNone of this is fatal to the idea. The derivation is coherent, the analysis is valuable, and the paper is transparent about its choices. The parameter-efficiency observation—a 400-neuron DS-STDP model beating a 900-neuron STDP model—is intriguing but suffers the same evaluation limitations. The broader claim that delays add representational capacity is plausible and partially supported by the STDP-with-zero-delay comparisons, but the evidence is not yet airtight.\n\nI would send this to peer review. The method and the analytical comparison deserve referee time. A serious revision should rerun the comparison with a fair DR-STDP configuration (or a hyperparameter sensitivity sweep), use validation-based early stopping, and repeat seeds. If the advantage disappears, the conclusion should be softened to “competitive.” As it stands, read the rule and the analysis; treat the MNIST ranking as provisional.","headline":"DS-STDP is a genuinely novel, clearly-derived extension of STDP for co-learning delays, but its headline 'consistent superiority' over DR-STDP is undermined by an unfair baseline and test-set oracle selection.","tokens_in":14593,"tokens_out":3853,"would_cite":true,"duration_ms":35470,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"The paper claims that DS-STDP, an STDP variant that shifts pre-synaptic spikes by a learned delay and updates delays with inverted kernels, consistently outperforms both DR-STDP and delay-free STDP across all tested network sizes on MNIST…","keywords":["spiking neural networks","synaptic delays","STDP","DS-STDP","unsupervised learning","MNIST","latency coding","neuromorphic computing"],"falsifier":"Re-run the same comparisons while matching DR-STDP's effective update rate to DS-STDP's, for example by scaling DR-STDP's trace amplitudes and time constants so that its average per-epoch weight and delay changes equal DS-STDP's, and check whether DS-STDP still wins at every network size; if DR-STDP catches up or surpasses under any such calibration, the claimed superiority is an artifact of the chosen DR-STDP hyperparameters.","tokens_in":13564,"feed_emoji":"🧠","tokens_out":6703,"duration_ms":87500,"temperature":0.7,"pith_summary":"The paper sets out to show that synaptic delays can be learned by the same kind of local, spike-timing-based rule that STDP uses for weights, and that doing so helps spiking networks classify. Its proposed rule, DS-STDP, shifts each pre-synaptic spike train by the synapse's current delay before applying the standard STDP weight update, and learns the delay itself with the same spike traces but with the potentiation and depression kernels swapped. The authors claim that, on the MNIST digit task with a widely used unsupervised spiking network, DS-STDP beats both DR-STDP, a prior delay-learning rule, and plain STDP at every network size from 100 to 900 neuron pairs, under both rate-based and latency-based readouts. If the claim holds, it gives neuromorphic and computational-neuroscience models a biologically plausible way to co-tune weights and delays, and it suggests that delay learning adds representational capacity that extra weights alone do not provide.","feed_headline":"Delay-shifted rule outperforms plain STDP on every network size","feed_subtitle":"Co-training synaptic weights and delays beats fixed-delay spiking models on digit classification.","key_machinery":"The central object is the delay-shifted STDP update. For a synapse with weight $w$ and delay $d$, the pre-synaptic spike train $S_{\\mathrm{pre}}(t)$ is replaced by the delayed train $S_{\\mathrm{pre}}(t-d)$ inside the weight update, so the rule sees spikes at the moment the post-synaptic neuron receives them. The delay update uses the same two spike traces but with the potentiation and depression kernels $K'_+$ and $K'_-$ swapped relative to the weight rule; in trace form it is $\\dot{d} = S_{\\mathrm{post}}(t) X'_{\\mathrm{pre}}(t-d) + S_{\\mathrm{pre}}(t-d) X'_{\\mathrm{post}}(t)$, where $X'$ denotes the corresponding spike trace. The paper shows that for isolated spike pairs this makes DS-STDP's initial updates numerically identical to DR-STDP's but anchored at the time of spike reception rather than spike generation, and that DS-STDP keeps its updates sparse because, in a discrete simulation, a delay increase can re-observe the same pre-synaptic spike at most twice unless the trace amplitude exceeds half a time step.","core_discovery":"On the paper's own terms, DS-STDP is a Hebbian rule that treats synaptic delay as a second learnable parameter. It updates weights with the standard STDP rule applied to delay-shifted pre-synaptic spikes, and it updates delays with the same pair of spike traces but with the potentiation and depression kernels flipped relative to the weight rule. The paper argues that this local, online rule makes a connection's delay decrease when the connection is potentiated and increase when it is depressed, matching the biological intuition that stronger connections act faster. It then claims that, in an unsupervised winner-take-all spiking network trained on MNIST, DS-STDP achieves the best test accuracy at every model size from 100 to 900 neuron pairs under both spike-rate and spike-responsiveness readouts, and that a 400-neuron DS-STDP network outperforms a 900-neuron delay-free network that has more trainable parameters.","pith_inferences":["A fairer comparison would sweep DR-STDP's hyperparameters: since the reported DR-STDP configuration halves trace amplitudes and time constants, DS-STDP's margin may partly reflect deliberate slowing of the baseline rather than a property of the rule itself.","The delay distributions learned by DS-STDP could be inspected directly: if delays encode input timing structure, synapses carrying early-arriving spikes should concentrate around informative pixels, an analysis the paper does not perform.","The claim that delays act like extra representational capacity suggests a testable equivalence: matching a delay-enabled network's accuracy with a pure-weight network may require more than doubling its width, not merely doubling parameter count.","Because DS-STDP is local and event-driven, it could be ported to online continual learning settings where non-local gradients are unavailable, though the paper does not test that setting."],"forward_implications":["If DS-STDP's advantage is real, unsupervised spiking networks can treat delay as a third dimension of synaptic state, and latency-based readouts should be preferred whenever delays are trained.","Synaptic delays can substitute for extra neurons: a 400-neuron DS-STDP network beats a 900-neuron STDP network with more parameters, which the paper interprets as a supra-additive increase in model capacity.","Because DS-STDP only shifts spikes by the current delay and updates on spike events, it can run online with a bounded buffer of length $d_{\\max}$, making it implementable in neuromorphic hardware.","The analytical comparison predicts that single-pair DS-STDP and DR-STDP updates have equal values but occur at different times, giving a concrete signature for distinguishing the two rules in simulations or recordings."],"supporting_citations":[{"why":"Supplies the unsupervised spiking network architecture and the training and evaluation procedure on which all three methods are tested.","marker":"[DC15]"},{"why":"Defines DR-STDP, the delay-learning baseline that DS-STDP is compared against and whose update timing is analyzed.","marker":"[NG24]"},{"why":"Provides the STDP integral formulation and exponential kernels that DS-STDP extends.","marker":"[KH00]"},{"why":"Provides the base-model implementation and the neuron, synapse, and weight-initialization hyperparameters used in the experiments.","marker":"[Haz+18]"},{"why":"Supplies the MNIST dataset used for all classification experiments.","marker":"[Lec+98]"},{"why":"Defines the cumulative and saturating spike traces that DS-STDP uses to replace convolved spike trains in discrete time.","marker":"[MDG08]"},{"why":"Supplies the power-law parameter dependence used to bound the synaptic weight updates during training.","marker":"[Güt+03]"},{"why":"Provides the comparison between weight-only and delay-only training that motivates treating delays as carrying representational capacity.","marker":"[GS23]"}],"fun_headline_variants":["DS-STDP learns synaptic delays to beat fixed-delay spiking nets","Extend STDP to learn delays: DS-STDP wins on digit classification","Co-train synaptic strength and delay in spiking networks for accuracy","Shifting spikes by learnable delays improves STDP classification","Learning delays with STDP: better digits without extra neurons"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The comparison assumes that DR-STDP's deliberately slowed training hyperparameters are a fair baseline; if the chosen settings underrepresent DR-STDP, DS-STDP's reported edge could come from calibration rather than from the rule.","fun_headline_variants_meta":{"raw":{"variants":["DS-STDP learns synaptic delays to beat fixed-delay spiking nets","Extend STDP to learn delays: DS-STDP wins on digit classification","Co-train synaptic strength and delay in spiking networks for accuracy","Shifting spikes by learnable delays improves STDP classification","Learning delays with STDP: better digits without extra neurons"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000375,"raw_usage":{"total_tokens":1970,"prompt_tokens":882,"completion_tokens":1088,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":498,"completion_tokens_details":{"reasoning_tokens":998}},"tokens_in":498,"tokens_out":1088,"duration_ms":8755,"temperature":1.0,"reasoning_tokens":998,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T00:09:10.156308+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Re-run the same comparisons while matching DR-STDP's effective update rate to DS-STDP's, for example by scaling DR-STDP's trace amplitudes and time constants so that its average per-epoch weight and delay changes equal DS-STDP's, and check whether DS-STDP still wins at every network size; if DR-STDP catches up or surpasses under any such calibration, the claimed superiority is an artifact of the chosen DR-STDP hyperparameters.","supporting_citations":[],"review_version":1}