Pith. sign in

REVIEW 3 major objections 5 minor 2 references

Extending Spike-Timing Dependent Plasticity to Learning Synaptic Delays

T0 review · 3 major / 5 minor · reviewed 2026-08-07 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read 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…

desk verdict 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. read the letter →

arxiv 2506.14984 v1 pith:UHOEHSRL submitted 2025-06-17 cs.NE cs.LG

Marissa Dominijanni , Alexander Ororbia , Kenneth W. Regan This is my paper · ORCID
classification cs.NEcs.LG
keywords spikingneuralnetworkssynapticdelaysSTDPDS-STDPunsupervisedlearningMNISTlatencycodingneuromorphiccomputing
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

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.

What carries the argument

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.

What would settle it

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.

Watch

Extended reading notes

Core claim

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.

Load-bearing premise

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.

Editorial extensions

If this is right

  • 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.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • 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.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

3 major / 5 minor

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.

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 (3)
  1. [Section 3.3.1, Table 4] 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.
  2. [Section 3.3.1, Tables 6-7] 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.
  3. [Section 3.3 and Section 4, Tables 6-8] 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.
minor comments (5)
  1. [Abstract] The phrase 'against another existing methods' should read 'against other existing methods'.
  2. [Section 2.3, Equations 19-20] 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.
  3. [Section 4, Figure 6] 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.
  4. [Section 4, Figure 7] 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.
  5. [Section 3.2, Equations 21-23] 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.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: DS-STDP is a new delay-shifted STDP rule and the empirical comparison is self-contained; the main evaluation caveats are experimental fairness, not definitional circularity.

full rationale

The paper's derivation chain is not circular. DS-STDP (Eqs. 9-12) is defined by shifting the pre-synaptic spike train by a learned delay and flipping the potentiative/depressive kernels for the delay update; these definitions do not presuppose the empirical outcome that DS-STDP outperforms DR-STDP or STDP. The sign convention for delay updates is borrowed from the external DR-STDP result [NG24], not from the present authors' prior work, and it is used as motivation rather than as a proof of the headline result. The only self-citations by author Ororbia ([Oro23], [OF24]) appear in background and concluding discussion and are not load-bearing. Evaluation concerns — Table 4 deliberately halves DR-STDP amplitudes/time constants, and Section 3.3.1 selects 'the best testing accuracy for a given epoch' — are threats to the fairness or stability of the empirical comparison, but they do not reduce any prediction to a fitted input by construction. No equations are equivalent to their inputs by definition, no fitted parameter is renamed as a prediction, and no uniqueness claim is imported from the authors' own prior work. The central claim therefore has independent empirical content, and any circularity score above 0 would be disproportionate.

Assumptions & free parameters 6 free parameters · 6 assumptions · 0 invented entities

The DS-STDP method depends on several hand-set hyperparameters (amplitudes, time constants, delay bound, normalization) and on standard domain assumptions about STDP, neuron models, and input encoding. The only assumption unique to this paper is the oracle-style test-set-based early stopping used in evaluation. No new physical entities are introduced.

free parameters (6)
  • Weight update amplitudes A+ and A- = A+ = 5e-4, A- = -5e-6
    Hand-set hyperparameters controlling DS-STDP weight potentiation and depression; also used for the STDP baseline. DR-STDP uses half these values in Table 4.
  • Delay update amplitudes A'+ and A'- = A'+ = 1.2e-4, A'- = -1.2e-2
    Hand-set amplitudes for delay updates; the asymmetry (depression amplitude much larger) directly controls how delays evolve and is central to DS-STDP's behavior.
  • Trace time constants tau+, tau-, tau'+, tau'- = 20 ms (DS-STDP/STDP), 10 ms (DR-STDP)
    Time constants set by hand; they define the STDP window and are different for DR-STDP to slow its updates.
  • Delay initialization and bound dmax = U(0,10) ms; dmax=10 ms
    Delays initialized uniformly and clamped to 10 ms; the maximum delay determines the buffering requirement and the delay shift scale.
  • Weight normalization constant wbar = 78.4
    Per-output L1 norm used in Eq. 28; taken from BindsNET implementation, affects effective weight scale.
  • Maximum spike rate nu_max = 127.5 Hz
    Poisson encoding peak rate; influences spike train statistics and learning dynamics.
assumptions (6)
  • domain assumption STDP is a valid Hebbian learning rule for unsupervised feature extraction in SNNs.
    The paper builds DS-STDP on the standard STDP formulation (Eq. 4) and relies on STDP's ability to train the Diehl-Cook network without labels.
  • domain assumption The relation that potentiating a synapse corresponds to decreasing its delay (from NG24).
    Section 2.2 adopts the observation 'potentiating synaptic connections corresponds to increasing weight values and decreasing delay values' [NG24] as the basis for the delay update sign.
  • domain assumption LIF/ALIF neuron models and delta synapses adequately capture spiking dynamics for the learning task.
    Section 3.3.2 uses ALIF for excitatory and LIF for inhibitory neurons with current-based delta synapses; accuracy of the model depends on these simulation choices.
  • domain assumption Poisson rate encoding with min-max scaling represents MNIST inputs adequately.
    Section 3.3 encodes pixel intensities as homogeneous Poisson spike trains without refractoriness; the learning signal depends on this encoding.
  • domain assumption Power-law weight dependence and post-hoc L1 normalization preserve STDP learning.
    Eq. 27 and Eq. 28 are applied after updates; the paper assumes these standard modifications do not distort the comparison.
  • ad hoc to paper The best test-set accuracy across training epochs is a valid model selection criterion.
    Section 3.3.1 states 'We use the best testing accuracy for a given epoch to simulate early halting under ideal conditions.' This is an oracle on the test set, not a realistic selection rule.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Extending Spike-Timing Dependent Plasticity to Learning Synaptic Delays." pith.science (2026). https://pith.science/paper/UHOEHSRL

@misc{pith2026250614984,
  author       = {Pith},
  title        = {Pith review of: Extending Spike-Timing Dependent Plasticity to Learning Synaptic Delays},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/UHOEHSRL}},
  note         = {Machine review of arXiv:2506.14984}
}
read the original abstract

Synaptic delays play a crucial role in biological neuronal networks, where their modulation has been observed in mammalian learning processes. In the realm of neuromorphic computing, although spiking neural networks (SNNs) aim to emulate biology more closely than traditional artificial neural networks do, synaptic delays are rarely incorporated into their simulation. We introduce a novel learning rule for simultaneously learning synaptic connection strengths and delays, by extending spike-timing dependent plasticity (STDP), a Hebbian method commonly used for learning synaptic weights. We validate our approach by extending a widely-used SNN model for classification trained with unsupervised learning. Then we demonstrate the effectiveness of our new method by comparing it against another existing methods for co-learning synaptic weights and delays as well as against STDP without synaptic delays. Results demonstrate that our proposed method consistently achieves superior performance across a variety of test scenarios. Furthermore, our experimental results yield insight into the interplay between synaptic efficacy and delay.

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

2 extracted references · 2 linked inside Pith

  1. [589]

    Learning represen- tations by back-propagating errors

    url: http://arxiv.org/abs/2311.09589 (visited on 06/08/2025). [Oro23] Alexander G. Ororbia. Brain-Inspired Machine Intelligence: A Survey of Neurobiologically- Plausible Credit Assignment. arXiv:2312.09257 [cs]. Dec. 2023. doi: 10.48550/ar Xiv . 2312 . 09257. url: http : / / arxiv . org / abs / 2312 . 09257(visited on 06/08/2025). [Ros24] Robert Rosenbaum...

    arXiv 2025

  2. [2236]

    Effect of geometrical irregularities on propagation delay in axonal trees

    doi: 10 . 1016 / S0166 - 2236(02 ) 02212 - 9. url: https : / / www . scie ncedirect . com / science / article / pii / S0166223602022129(visited on 03/12/2025). [MKS91] Y. Manor, C. Koch, and I. Segev. “Effect of geometrical irregularities on propagation delay in axonal trees”. In: Biophysical Journal 60.6 (Dec. 1991), pp. 1424–1437. issn: 0006-3495. doi: ...

    arXiv 2025

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.