{"id":"d1aeac92-f32b-48bc-a8ef-3f4887aedbff","arxiv_id":"1908.08655","paper_version":3,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":7,"one_line_summary":"A spiking predictive-coding network with a local, spike-triggered alignment rule achieves 4.72% error on MNIST in a single online pass and forgets less on split continual-learning benchmarks.","lead":"This paper builds a spiking neural network that learns by predicting its own inputs and adjusting local connections when predictions fail, and reports single-pass image classification accuracy competitive with spiking networks trained for many epochs. It matters because the learning rule is local and event-driven, making it a candidate for low-power neuromorphic hardware and online continual learning systems.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"ST-LRA update is internally inconsistent as written: Eq. 2 and Eq. 5 disagree on which error signal updates W_l, and the stated weight application swaps ΔW and ΔE, leaving the central learning rule underspecified.","rationale":"The reader identified the sufficiency of ST-LRA credit assignment as the weakest assumption. I agree with that general area, but the stress-test reveals a sharper, more specific problem: the learning rule is not just under-justified, it is internally inconsistent as written. Eq. 2 and Eq. 5 cannot both be true under a single consistent indexing convention, and the update application line conflicts with the definitions of ΔW and ΔE. Since the paper provides no code, the published equations are the only specification of the algorithm, so this inconsistency directly blocks verification of the central claim: that a local spiking predictive-coding rule trains the network to competitive MNIST accuracy and reduced forgetting. This does not prove the underlying idea is wrong. The figure, the transpose relationship ΔE = −β(ΔW)^T, and the surrounding text all suggest a plausible intended LRA rule, and the empirical results may well be reproducible once the indexing and sign conventions are fixed. But a reader cannot determine that from the manuscript. The paper's own limitations section is honest about biological simplifications, and the authors explicitly disclaim a full solution to catastrophic forgetting, which is to their credit. Those issues are not the problem. The problem is that the central algorithm, as specified, is ambiguous enough that the reported results cannot be traced to the described mechanism. This reinforces the reader's CONDITIONAL verdict: the paper needs a corrected, self-consistent equation set, explicit tensor shapes, and ideally released code, before the empirical claims can be fully evaluated. I do not recommend moving to REJECT because the intended algorithm appears recoverable from context, but the conditional status is clearly warranted.","tokens_in":60572,"tokens_out":8957,"duration_ms":88284,"concrete_test":"Implement Algorithm 1 exactly as written in the paper (Eqs. 2, 3, 5, and the stated application W−α_uΔE, E−α_uΔW) using the reported hyperparameters and training protocol, and run MNIST. If test error reproduces 4.72±0.11%, the inconsistency is typographical and the mechanism is sound; if training diverges or error is far worse, the empirical claims depend on an unstated correction. A complementary check is to swap in the intended predictive-coding variant (W_l updated by e_l·s_l^T, W−α_uΔW, E−α_uΔE) and verify that it reproduces Table 1.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim that local ST-LRA trains a deep spiking predictive-coding network depends on the correctness and reproducibility of Eqs. 2 and 5. As printed, those equations are inconsistent. Eq. 2 defines e_l(t) = W_l·s_l(t) − z_l(t), so W_l predicts layer l's own filtered activity and the natural predictive-coding update would be proportional to e_l(t)·s_l(t)^T. Eq. 5 instead updates ΔW_l = e_{l−1}(t)·(s_l(t))^T, which only makes sense if Eq. 2 is off by one, i.e., W_l predicts z_{l−1}(t) from s_l(t). Figure 1 and the use of E_l·e_{l−1}(t) in Eq. 3 suggest the off-by-one reading is intended, but the paper never states it and Algorithm 1 repeats the same indexing. The subsequent line, \"W_l ← W_l − α_u ΔE_l and E_l ← E_l − α_u ΔW_l,\" additionally swaps the two update matrices relative to their definitions; dimensions only work out under the alternate indexing and with a specific convention for W_l and E_l shapes, which is not given. Without shipped code, a reader cannot tell which variant produced Table 1 or Table 4. If the published equations are implemented literally, the learning rule may not perform credit assignment at all, in which case the reported accuracy and continual-learning results cannot be attributed to the described mechanism. This is more concrete than a general absence of convergence analysis: the update rule as specified is not fully determined.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes the Spiking Neural Coding Network (SpNCN), a recurrent network of leaky integrate-and-fire neurons trained by a spike-triggered local representation alignment (ST-LRA) update derived from predictive coding. The model is evaluated online and in continuous time on a bouncing-ball prediction task, on MNIST, Fashion-MNIST, Stanford OCR, and Caltech-101 classification, on semi-supervised variants of MNIST/Fashion-MNIST, and on Split-MNIST/NotMNIST continual-learning benchmarks. The central claims are that ST-LRA provides a local, backprop-free credit-assignment rule for multi-layer spiking networks, that SpNCN is competitive with existing SNNs after a single pass over the data, that it can exploit unlabeled data, and that its sparse spike-based representations forget less when tasks are presented sequentially.","tokens_in":60889,"tokens_out":9388,"duration_ms":98658,"significance":"If the learning mechanism performs as claimed, the paper offers a genuinely local, online, spike-based alternative to backpropagation-style training for spiking networks and a concrete test of the hypothesis that sparsity mitigates catastrophic forgetting. The paper has real strengths: all classification results are evaluated on held-out test sets rather than training data; the authors implement and compare against derivative-free spiking baselines under the same LIF SRM; the sparse weight-update counts in the bouncing-ball experiment give a concrete computational-economy datum; and the semi-supervised comparison is an informative stress test. The significance is conditional, however, because the central learning rule is not consistently specified and key simulation hyperparameters are missing.","major_comments":[{"comment":"The learning rule as printed is not internally consistent. Eq. (2) defines e_l(t) = W_l s_l(t) − z_l(t), so W_l appears to predict layer l's own filtered activity from its own spikes, whereas Eq. (5) updates W_l with ΔW_l = e_{l−1}(t)(s_l(t))^T, which is a predictive-coding update only under an off-by-one indexing convention in which W_l predicts z_{l−1}(t) from s_l(t). The subsequent sentence \"W_l ← W_l − α_u ΔE_l and E_l ← E_l − α_u ΔW_l\" then swaps the two matrices relative to their definitions, and the shapes of W_l and E_l are never stated. Algorithm 1 repeats Eqs. (2) and (5) without resolving the indexing. Since no code is provided, a reader cannot determine which variant produced Tables 1–4; if the equations are implemented literally, the rule may not be performing predictive-coding credit assignment at all. Please restate the model with consistent subscripts, specify the dimensions of all weight matrices, and reconcile Algorithm 1 with the equations.","section":"Section 2.2, Eqs. (2) and (5), Algorithm 1"},{"comment":"The firing threshold v_thr is never reported. The text states only that a threshold is chosen and that voltages operate in the [0,1] decivolt range, while the MNIST experiments report 4 layers of 1000 LIF units, T_st = 100 ms, Δt = 0.25 ms, α = 0.0025, β = 1.0, and K = 63.75 Hz, but not the threshold, the membrane time constant τ_m, R_m, γ_m, τ_f, or whether an inter-stimulus interval T_ist was used. Because the threshold controls spiking frequency, the sparse spike-triggered updates, and hence the online learning dynamics, the headline 4.72% single-pass error cannot be reproduced or checked against the claim that sparse activity drives learning. Please provide a complete hyperparameter table for every experiment, including the bouncing-ball and continual-learning settings.","section":"Sections 2.2 and 3.2"},{"comment":"The reported standard errors (±0.001% to ±0.003% over 10 trials) are implausibly small. For a binary accuracy near 0.76 on a held-out test set of even 10,000 examples, the per-trial standard deviation is roughly 0.4 percentage points and the standard error of the mean is about 0.13 percentage points, so the printed error bars are one to two orders of magnitude too small. As written, the error bars imply either that the metric is not what is stated, a typo, or an averaging artifact. This matters because the continual-learning advantage over the SNN baseline is the main quantitative support for the \"less forgetting\" claim, and its precision must be reported correctly.","section":"Table 4"},{"comment":"The bouncing-ball result is a single run: one stream of K = 2000 frames, one simulated 30-second sequence, and no repeated seeds, initializations, or error bars, and the 1000-frame train/freeze boundary is not justified. The Frame(t−1) comparison is useful, but a single-run pSE difference (6.672 vs 10.225) is anecdotal evidence for the model's predictive-tracking ability, which is one of the paper's three experimental pillars. Please report results over multiple random streams and initializations, with variance or interquartile ranges.","section":"Section 3.1, Figure 3"},{"comment":"The continual-learning experiments report only aggregate average accuracy (ACC) after the full stream. Aggregate accuracy does not directly measure forgetting; a model can achieve high ACC while still overwriting earlier tasks if later tasks are easier or if the final average is dominated by later performance. Since the stated goal is to determine whether spiking sparsity reduces forgetting, please also report per-task accuracy after each task, backward transfer, or an explicit forgetting measure such as final average per-task accuracy versus peak per-task accuracy.","section":"Table 4 and Section 3 (\"On Catastrophic Forgetting\")"}],"minor_comments":[{"comment":"The first form of the trace filter writes z_l(t) on both sides; please use an explicit time index such as z_l(t+Δt) to remove the ambiguity between an in-place recurrence and a fixed-point equation.","section":"Equation (1)"},{"comment":"The caption says the columns vary \"the proportion of samples that arrive labeled,\" but the first column is 0% and is described as fully supervised; please clarify whether the column percentages refer to labeled or unlabeled samples.","section":"Table 3"},{"comment":"Table 4 labels the baseline \"SNN df-BDA,\" which should read \"SNN df-BFA,\" and Section 3.1 plus Figure 3 use \"SpTNCN\" where \"SpNCN\" is meant.","section":"Table 4 and Section 3.1"},{"comment":"The caption says performance was \"averaged over 10 trails\"; this should be \"10 trials.\"","section":"Section 3.2, Table 1 caption"},{"comment":"No hyperparameters are reported for the df-BFA and df-DRTP baselines (learning rates, numbers of layers and units, feedback weight initialization, training epochs, tuning procedure). Please provide these settings so the comparisons in Tables 1–3 cannot be attributed to under-tuned baselines.","section":"Appendix (df-BFA and df-DRTP)"},{"comment":"In COMPUTE STATES, currents and voltages are updated layer by layer before the prediction/error loop; please state explicitly whether e_l(t) is computed from the post-update z_l(t) or from the previous value, since this timing changes the effective ST-LRA update.","section":"Algorithm 1"}],"recommendation":"major_revision","confidential_remarks":"The paper is labeled \"Preprint, Work in Progress,\" and the most consequential defect is the underspecified ST-LRA rule. The off-by-one indexing and the swapped update matrices are plausibly typographical rather than fundamental, but without a corrected, dimensionally consistent pseudocode or released code, Sections 3–5 cannot be audited. I would make code release or a complete algorithmic appendix a condition of acceptance. The self-citation pattern is notable but not disqualifying; the empirical comparisons would be more persuasive if at least one externally implemented spiking baseline were available."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The paper proposes a spiking version of neural predictive coding (SpNCN) with a local, event-driven learning rule (ST-LRA) and claims competitive single-pass MNIST error (4.72%), decent semi-supervised behavior, and reduced forgetting on Split MNIST/NotMNIST. The idea is a legitimate extension of Ororbia's earlier LRA work, and the MNIST result, if real, would be notable for a fully online, backprop-free spiking network.\n\nWhat the paper does well: the spiking reformulation is clearly described at a high level, the ST-LRA rule is genuinely local and hardware-friendly, and the experiments are wide enough to test the model in several regimes. The author is also upfront about biological limitations (e.g., the lack of separate excitatory/inhibitory synapses) and even cautions that natural sparsity alone won't solve catastrophic forgetting. That honesty helps.\n\nThe soft spots are not minor. The central equations are internally inconsistent. Eq. 2 defines e_l = W_l s_l − z_l, so W_l predicts layer l's own trace. Eq. 5 updates W_l with e_{l−1} s_l^T, which only makes sense if W_l is predicting layer l−1. The subsequent line updates W_l with ∆E_l and E_l with ∆W_l, compounding the confusion. The paper never states the intended indexing or matrix shapes, and Algorithm 1 repeats the same pattern. With no code shipped, a reader cannot tell which variant produced the reported tables. This is more than a missing proof of convergence; the learning rule as printed is not fully determined.\n\nOther issues: v_thr and some other hyperparameters are never specified. The bouncing-ball experiment is a single run. The baselines (df-BFA, df-DRTP) are self-implemented, which is normal, but they are not standard off-the-shelf baselines. Table 4's error bars are implausible: standard errors of ±0.001–0.003% over ten trials would require almost zero variance. Those numbers look like a reporting mistake, but they undermine trust.\n\nBottom line: the conceptual framework has value for the neuromorphic and spiking-network community, and the empirical claims are moderate and mostly plausible. But the paper in its current form cannot be independently reproduced or even fully interpreted. A serious referee should see it, but the author needs to correct the equations, release code, and fix the error statistics before publication. I'd talk to them if I could; the idea is worth one more round.","headline":"A spiking predictive-coding network with a promising local rule, but the central equations are inconsistent and the empirics are hard to verify.","tokens_in":61434,"tokens_out":5965,"would_cite":false,"duration_ms":58124,"reading_group":"maybe","serious_thinker":"no","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"This paper claims that a spiking predictive-coding network with purely local spike-triggered weight updates can learn online from one-pass data streams, matching multi-epoch spiking networks on MNIST and forgetting less on sequential…","keywords":["spiking neural networks","predictive coding","continual learning","online learning","semi-supervised learning","local representation alignment","leaky integrate-and-fire","catastrophic forgetting"],"falsifier":"Train the reported four-layer SpNCN on MNIST with the error signals $e_\\ell(t)$ replaced by independent noise of the same magnitude during learning, so that ST-LRA still updates but carries no predictive information: if test error stays near $4.72\\%$, the mismatch computation is not the source of learning, and if it collapses toward chance, the predictive-coding error is confirmed as the mechanism.","tokens_in":60318,"feed_emoji":"🧠","tokens_out":17636,"duration_ms":155966,"temperature":0.7,"pith_summary":"The paper sets out to establish that the iterative guess-and-check loop of neural predictive coding can be rebuilt with spiking neurons and still learn useful representations from one-pass data streams, without backpropagation or even spike-timing-dependent plasticity. The model predicts each layer's filtered spike activity, computes mismatch signals in dedicated error neurons, and adjusts synapses by a local outer-product rule whenever a spike arrives. If correct, spiking networks gain an online, biologically plausible learning rule that suits neuromorphic hardware, and the sparse spike codes appear to reduce catastrophic forgetting on sequential tasks. The reported evidence includes $4.72\\%$ MNIST test error after a single pass, better average accuracy than its baselines on Split MNIST and NotMNIST, and usable semi-supervised learning as labels become rare.","feed_headline":"One-pass spiking net hits 4.72% MNIST error","feed_subtitle":"Local spike-triggered updates replace backprop; same network handles semi-supervised and continual-learning streams.","key_machinery":"The mechanism that carries the argument is the spiking predictive-coding loop closed by dedicated error neurons and a spike-triggered local update rule. Predictions $z^\\mu_\\ell = W_\\ell s_\\ell(t)$ are compared against low-pass filtered spike traces $z_\\ell(t)$ to produce error signals $e_\\ell(t) = z^\\mu_\\ell - z_\\ell(t)$; these errors are fed back through error synapses $E_\\ell$ into the recurrent current equation $J_\\ell(t) = (1-\\kappa)J_\\ell(t) + \\kappa(-\\gamma_J J_\\ell(t) + \\varphi(-e_\\ell(t) + E_\\ell e_{\\ell-1}(t)))$ for intermediate layers, with only $E_L e_{L-1}(t)$ for the top layer. Learning is by Spike-Triggered Local Representation Alignment (ST-LRA), the event-driven rule $\\Delta W_\\ell = e_{\\ell-1}(t) s_\\ell(t)^\\top$, $\\Delta E_\\ell = -\\beta s_\\ell(t) e_{\\ell-1}(t)^\\top$, an error-driven Hebbian outer-product update analogous to the delta rule. The paper uses a leaky integrate-and-fire spike-response model, but states the same three computations would accommodate richer neuron models such as Izhikevich or Hodgkin-Huxley units.","core_discovery":"The central discovery claimed is that a spiking network built on predictive coding can learn online without a global error signal or repeated exposure to data. At every simulated time step, a layer $\\ell$ predicts the filtered spike trace of another population, $z^\\mu_\\ell = W_\\ell s_\\ell(t)$, error neurons form the mismatch $e_\\ell(t) = z^\\mu_\\ell - z_\\ell(t)$, and these error activities are routed through error synapses $E_\\ell$ back into the membrane currents $J_\\ell(t)$ that drive the spike-response model. Synaptic change happens only when spikes occur, under the Spike-Triggered Local Representation Alignment rule $\\Delta W_\\ell = e_{\\ell-1}(t) s_\\ell(t)^\\top$ and $\\Delta E_\\ell = -\\beta s_\\ell(t) e_{\\ell-1}(t)^\\top$. The paper reports that with four layers of leaky integrate-and-fire units, this loop reaches $4.72\\%$ MNIST test error in one online pass, predicts a bouncing-ball video stream with lower error than a frame-repetition baseline, keeps working when most labels are missing, and retains more accuracy across Split MNIST and NotMNIST task streams than an equivalently sized backprop-trained ANN or a spiking network trained with derivative-free broadcast feedback alignment.","pith_inferences":["Inference: the claimed continual-learning gains are moderate, and the authors themselves note that sparsity alone will not solve catastrophic forgetting; a direct next test implied by their results is to combine ST-LRA with a complementary consolidation mechanism and measure forgetting on the same Split MNIST stream.","Inference: the generative side of the model suggests the same spike-driven loop could be used for sequence prediction and motor control, not just classification, by treating prediction error on future sensor frames as the learning signal in an event-camera or robotics stream.","Inference: because the paper does not compare its spiking version against a rate-coded version with the same local rule, a useful test is to run the same architecture with continuous units and non-spiking local representation alignment on the one-pass benchmarks to see whether spiking itself, rather than the update rule, drives the reported robustness."],"forward_implications":["A four-layer SpNCN reaches $4.72 \\pm 0.11\\%$ test error on MNIST after one online pass, placing it alongside spiking networks that train over many epochs and above the paper's implemented spiking baselines.","On the continual-learning streams Split MNIST and NotMNIST, the SpNCN achieves $76.455\\%$ and $77.945\\%$ average accuracy, outperforming the self-implemented spiking baseline and a backprop-trained ANN under task-boundary fuzzing.","With labels on only $0.5\\%$ of stream samples, a two-layer SpNCN still gets $24.08\\%$ MNIST error versus $37.35\\%$ for the derivative-free broadcast feedback alignment baseline, showing the generative side makes unlabeled data usable.","On the bouncing-ball stream, the SpNCN's prequential squared error reached 6.672 versus 10.225 for a frame-repetition baseline, and it made only about 46,393 lower-layer and 28,998 upper-layer weight updates over 300,000 simulation steps.","Because ST-LRA is local and event-driven, the paper expects it to combine with STDP and to fit neuromorphic hardware where sparse spike-driven updates translate to energy savings."],"supporting_citations":[{"why":"This supplies the non-spiking neural predictive coding model whose framework the SpNCN extends to spiking units.","marker":"[73]"},{"why":"This introduces the discrepancy-reduction family of learning rules from which spike-triggered local representation alignment is adapted.","marker":"[74]"},{"why":"This provides the local representation alignment update rule and its biological motivation as an error-driven Hebbian rule.","marker":"[75]"},{"why":"This grounds the error-neuron and top-down prediction architecture in the free-energy formulation of predictive coding.","marker":"[29]"},{"why":"This defines the spike-train communication paradigm and the spiking neuron modeling framework the system builds on.","marker":"[56]"},{"why":"This supplies the canonical predictive-coding circuitry with separate prediction and error neurons that motivates the SpNCN's architecture.","marker":"[88]"},{"why":"This provides the STDP-trained MNIST baseline and the Poisson spike-train encoding used in the image experiments.","marker":"[19]"},{"why":"This provides the derivative-free broadcast feedback alignment algorithm whose spiking implementation is the paper's primary comparison baseline.","marker":"[91]"}],"fun_headline_variants":["Spiking predictive coding learns streams without backprop","One-pass spiking net cuts forgetting on task streams","Spike-triggered local updates rival backprop on MNIST","Biologically plausible net learns continually from data streams","Forget-free learning from video and task streams"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The whole approach depends on the assumption that the mismatch signals computed at each layer, together with the local spike-triggered weight updates, are enough to train every layer correctly, a premise the paper motivates by analogy to the delta rule but does not prove or isolate.","fun_headline_variants_meta":{"raw":{"variants":["Spiking predictive coding learns streams without backprop","One-pass spiking net cuts forgetting on task streams","Spike-triggered local updates rival backprop on MNIST","Biologically plausible net learns continually from data streams","Forget-free learning from video and task streams"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001033,"raw_usage":{"total_tokens":4395,"prompt_tokens":1038,"completion_tokens":3357,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":654,"completion_tokens_details":{"reasoning_tokens":3280}},"tokens_in":654,"tokens_out":3357,"duration_ms":23595,"temperature":1.0,"reasoning_tokens":3280,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T11:33:17.836437+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Train the reported four-layer SpNCN on MNIST with the error signals $e_\\ell(t)$ replaced by independent noise of the same magnitude during learning, so that ST-LRA still updates but carries no predictive information: if test error stays near $4.72\\%$, the mismatch computation is not the source of learning, and if it collapses toward chance, the predictive-coding error is confirmed as the mechanism.","supporting_citations":[{"cited_title":"Lifelong Neural Predictive Coding: Learning Cumulatively Online without Forgetting","cited_arxiv_id":"1905.10696","evidence_quote":"This supplies the non-spiking neural predictive coding model whose framework the SpNCN extends to spiking units."},{"cited_title":"G., AND MALI , A","cited_arxiv_id":null,"evidence_quote":"This provides the local representation alignment update rule and its biological motivation as an error-driven Hebbian rule."},{"cited_title":"Predictive coding under the free-energy principle","cited_arxiv_id":null,"evidence_quote":"This grounds the error-neuron and top-down prediction architecture in the free-energy formulation of predictive coding."},{"cited_title":"Networks of spiking neurons: the third generation of neural network models","cited_arxiv_id":null,"evidence_quote":"This defines the spike-train communication paradigm and the spiking neuron modeling framework the system builds on."},{"cited_title":"P., AND BALLARD , D","cited_arxiv_id":null,"evidence_quote":"This supplies the canonical predictive-coding circuitry with separate prediction and error neurons that motivates the SpNCN's architecture."},{"cited_title":"P., AND TWEED , D","cited_arxiv_id":null,"evidence_quote":"This provides the derivative-free broadcast feedback alignment algorithm whose spiking implementation is the paper's primary comparison baseline."}],"review_version":1}