{"id":"252679d7-acd3-4a0f-92f8-c12efb7e23c2","arxiv_id":"2607.16027","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"A constrained Hebbian rule produces audiovisual representations with lower task-information cost (retained input information per unit of task-relevant information) than sparse backpropagation and DDTP at comparable accuracy in several settings.","lead":"An excitatory Hebbian learning rule, which strengthens synapses when neurons fire together, produces audiovisual representations with lower task-information cost than sparse backpropagation or DDTP at similar accuracy in several settings. The result supports viewing Hebbian plasticity as synaptic resource allocation rather than as a general accuracy booster.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"CTI advantage may reflect VIB bound-tightness differences rather than learning-rule efficiency; a matched-estimator check is needed.","rationale":"The reader's weakest assumption correctly identifies the load-bearing issue: the paper's empirical conclusion is entirely about CTI ordering across learning rules, and CTI is computed as a ratio of a variational upper bound to a variational lower bound. My own reading confirms that the paper is carefully hedged and the paired statistics are appropriate for the numbers as reported, but the statistical tests validate those numbers, not the mapping from them to true information costs. The specific preprocessing asymmetry — per-layer z-scoring and [0,1] rescaling for Hebbian/nonnegative BP versus tanh for BP/DDTP — gives a concrete reason the bound gaps could differ systematically. The fact that shallow nonnegative BP with the same preprocessing reaches the same low-CTI regime supports this concern, although deep nonnegative BP collapses, so preprocessing alone cannot explain the full pattern. The proposed whitening and gap-reporting check would settle whether the CTI advantage survives matched input scaling and comparable bound tightness. Because the outcome is genuinely open, CONDITIONAL is the appropriate verdict: the claim is plausible and well-hedged, but the central metric needs this robustness check before acceptance.","tokens_in":29799,"tokens_out":7089,"duration_ms":78131,"concrete_test":"Take the deep VGGSound100 bimodal frozen representations underlying Table 2 and re-run Phase-2 VIB evaluation after whitening each h (zero mean, unit variance per feature) for all learning rules, keeping the same VIB architecture, K=256, and beta=1e-2. Also report the bound-gap term KL[q(z)||N(0,I)] for each method. If Hebbian still shows CTI near 22 while BP/DDTP remain near 190/100 and the gap terms are comparable in size, the claim is robust. If the ordering shrinks or reverses, the reported CTI advantage is an input-scaling/bound-tightness artifact rather than a property of the Hebbian rule.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The central claim rests on CTI = I(Z;H)/I(Z;Y) in Eq. 9. Section 3.2 estimates I(Z;H) as E_h KL[q(z|h)||N(0,I)], which is an upper bound: it equals I(Z;H) + KL[q(z)||N(0,I)] (Eq. 4). Section 3.2 also estimates I(Z;Y) via a variational lower bound. The ratio of an upper bound to a lower bound is not an unbiased estimate of the true ratio, and the two gaps can differ across learning rules. There is a concrete mechanism: Hebbian and nonnegative-BP hidden activations are Z-score normalized and min-max rescaled to [0,1] at every layer (Secs. 3.1, 3.5), while BP/DDTP use tanh activations with no per-layer rescaling. Because the VIB encoder is a single linear layer into diagonal-Gaussian parameters, the scale and distribution of its input h directly affect how tightly a diagonal Gaussian posterior can fit while preserving classification accuracy. Thus the reported Hebbian advantage in Table 1 (geometric CTI ratios 5.37 and 6.73) may be inflated by systematically smaller bound gaps for Hebbian representations rather than by genuinely lower information cost. The paper's own shallow nonnegative-BP control, which uses the same preprocessing and also achieves low CTI, shows that preprocessing correlates with low CTI, though deep nonnegative BP collapses. The claim therefore needs a matched-estimator or bound-tightness check before the learning-rule interpretation is secure.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper tests whether a strictly local, excitatory, competitive Hebbian rule (a variant of Oja's PCA rule with nonnegativity and normalization) allocates representational resources more efficiently than backpropagation (BP) and Dense Difference Target Propagation (DDTP) under matched sparsity and architectural constraints. Using fixed VideoMAE/AudioMAE embeddings from AVE, Kinetics-Sounds, and VGGSound100, the authors train shallow and deep MLPs with each rule, freeze the encoder, and fit a post hoc Variational Information Bottleneck (VIB) module. The central metric is Task-Information Cost, CTI = I(Z;H)/I(Z;Y), estimated from VIB bounds. The main empirical claim is that Hebbian networks achieve lower CTI than sparse BP and DDTP in compressed comparisons, with accuracy remaining comparable in several settings, and that this supports interpreting Hebbian plasticity as a resource-allocation mechanism rather than a general accuracy-maximizing strategy.","tokens_in":30214,"tokens_out":3012,"duration_ms":32563,"significance":"If the central claim holds, the paper would provide a valuable demonstration that a biologically local Hebbian rule can shift the information-efficiency trade-off in a controlled assay, with implications for synaptic resource allocation theories. Strengths include the two-stage protocol that separates representation learning from the information-theoretic readout, matched sparsity and architecture across rules, paired seed-level statistical testing with Holm correction, and a set of ablations (MNIST, GHA, classical Oja, PCA-to-readout, depth-scaling). The manuscript is also unusually candid about its limitations, including the instability of deep nonnegative BP and the 10-layer Hebbian control. However, the central quantitative claim rests on a CTI estimate whose numerator and denominator are variational bounds of opposite direction; if bound tightness differs systematically across learning rules, the reported Hebbian advantage could be a measurement artifact rather than a property of the learning rule. The significance is therefore contingent on a matched-estimator or bound-tightness check.","major_comments":[{"comment":"CTI is defined as I(Z;H)/I(Z;Y), but the implementation estimates I(Z;H) by E_h KL[q(z|h)||r(z)] = I(Z;H) + KL[q(z)||r(z)] (Eq. 4, an upper bound) and I(Z;Y) by a variational lower bound. The ratio of an upper bound to a lower bound is not a bound on the true ratio, and the two gaps can differ across learning rules. The paper's own preprocessing creates a concrete mechanism: Hebbian and nonnegative-BP activations are Z-score normalized and min-max rescaled to [0,1] (Secs. 3.1, 3.5), while BP/DDTP use tanh activations without rescaling. Since the VIB encoder is a single linear layer into diagonal-Gaussian parameters, the scale and distribution of h directly affect how tightly a diagonal Gaussian posterior can fit. Thus the large Hebbian advantage in Table 1 (geometric CTI ratios 5.37 and 6.73) may reflect smaller bound gaps for Hebbian representations rather than genuinely lower informati","section":"§3.2 and §3.6.1, Eq. (9)"},{"comment":"The main tabular comparison is selected at β=10^-2, and the text states that higher-β operating points for DDTP and BP were excluded because they show a rapid decline in performance. Since CTI generally decreases as β increases, excluding the higher-β points of the reference methods removes exactly the operating points where BP/DDTP might achieve lower CTI. To support the claim that Hebbian learning 'achieves lower CTI than sparse BP and DDTP,' the authors should report complete β trajectories and compare operating points on a comparable basis (e.g., matched I(Z;Y) or matched Top-1 accuracy), not only at a fixed β chosen post hoc. Without this, the reported advantage may be an artifact of the selected operating point.","section":"§4.1.1, Fig. 2 and Table 1"},{"comment":"The paired Wilcoxon analysis (N=30, pHolm=3.73e-9) is correctly applied to log-transformed CTI ratios and supports the claim that the Table 1 values differ in the matched conditions. However, the analysis inherits the validity of the CTI estimator. If the bound-tightness concern in the first major comment is not addressed, the statistical test only shows that the estimated CTI values differ, not that the true information costs differ. The authors should either defend the comparability of the bounds more rigorously or report the statistical test on a corrected estimator. This is not a call to remove the statistics, but to ensure the quantity being tested is the quantity of scientific interest.","section":"§3.6.1, statistical analysis"}],"minor_comments":[{"comment":"The activation function φ(a_j) is defined as a min-max rescaling over 'the corresponding hidden layer,' but it is not stated whether the min/max are computed per batch, per dataset, or over a running statistic. Since the VIB input distribution depends on this, please clarify the exact normalization procedure in the main text.","section":"§3.1, Eq. (5)"},{"comment":"The row 'BP (nonneg.) Deep*' contains only em-dashes, and the footnote says dense connectivity only for the first BP row. This is confusing. If deep nonnegative BP was not trained on AVE/Kinetics-Sounds and only a single run exists on VGGSound100, state this explicitly in a table note rather than using an empty row.","section":"Table 1"},{"comment":"The 10-hidden-layer depth-scaling control reports an approximate Top-1 range of 7–10% with no multi-seed estimate. Since this is interpreted as a limitation, it would be helpful to state the chance level for the 100-class VGGSound100 split (1%) so readers can assess how close to chance that range is.","section":"§4.1.4, Table 4"},{"comment":"There are several typographical issues: 'T raining Algorithm' in Section 3.5.1, 'A VE' with a space throughout, and inconsistent use of 'nonneg.' vs. 'nonnegativity-constrained' in tables. These do not affect the science but should be cleaned up.","section":"Global"}],"recommendation":"major_revision","confidential_remarks":"The central concern is not circularity—the representations are learned independently of the VIB readout, which is a genuine strength. The issue is the comparability of the CTI estimator across learning rules. The matched-estimator check I request is well within the scope of the paper and could be added as a control experiment. I would also encourage the editor to consider that the paper relies heavily on the same group's prior Hebbian work (e.g., [22,23]); this is not a reason to reject, but reviewers should be alert to the possibility of confirmation bias in the choice of compared operating points."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: this is a genuinely new empirical result wrapped in a metric that may be biased in the paper's own favor. The Hebbian update itself comes from earlier work, but the controlled audiovisual comparison, matched sparsity, seed-level statistics, and the CTI metric are new. The paper is honest about its scope.\n\nWhat it does well: fixed pretrained embeddings isolate downstream plasticity; dense-to-sparse pruning is biologically motivated; the MNIST raw-input control and the shallow nonnegative-BP comparison are exactly the right checks; the statistics are paired and corrected; the authors openly report the deep nonneg BP collapse and the 10-layer instability. Credit where due.\n\nThe soft spot is the CTI ratio. Eq. (9) divides an upper-bound estimate of I(Z;H) by a lower-bound estimate of I(Z;Y), so the ratio is not a bound, and its tightness can differ across rules. The concrete mechanism in the stress test looks plausible: Hebbian and nonnegative-BP activations are z-scored and rescaled to [0,1], while BP and DDTP use tanh without per-layer rescaling. A diagonal-Gaussian VIB posterior will fit the rescaled inputs differently, potentially lowering estimated I(Z;H) for Hebbian for reasons unrelated to the learning rule. The fact that shallow nonnegative BP—which uses the same preprocessing—also lands in the low-CTI regime is a red flag, not a refutation, but it means the headline comparison needs a matched-estimator check.\n\nOther issues are minor by comparison: the main table is beta=10^-2, higher-beta points for BP/DDTP are excluded because performance collapses, and deep nonneg BP is missing, so the deep comparison lacks that control. These are addressable.\n\nBottom line: the resource-allocation interpretation of Hebbian plasticity is plausible and the framework is useful, but the central quantitative claim is provisional until the bound-tightness question is settled. I would send this to peer review and ask for diagnostics: bound-gap estimates for each rule (e.g., aggregated-posterior KL to the prior), all beta trajectories, and code/data. With those, the CTI comparison becomes interpretable. This is a serious paper for computational-neuroscience readers, not an audiovisual benchmark.","headline":"New empirical comparison with a plausible confound in the CTI metric; the resource-allocation story is provisional until bound-tightness is checked, but the paper deserves a serious referee.","tokens_in":30691,"tokens_out":2230,"would_cite":true,"duration_ms":22131,"reading_group":"maybe","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 a strictly local, excitatory Hebbian learning rule, evaluated as a synaptic resource-allocation mechanism, produces representations with lower task-information cost than sparse backpropagation and target propagation at","keywords":["Hebbian learning","synaptic resource allocation","variational information bottleneck","task-information cost","sparse neural networks","nonnegative weights","auditory-visual learning","local plasticity"],"falsifier":"Take the three frozen representation sets (Hebbian, sparse BP, DDTP) for one dataset and architecture, and compute I(Z;H) not with the KL-to-N(0,I) proxy but with the aggregated-posterior KL or a nonparametric estimator (e.g., k-NN) on samples from the VIB encoder, then recompute CTI at matched β and accuracy. If the Hebbian CTI advantage over BP/DDTP vanishes or reverses, the paper's resource-allocation conclusion is an estimation artifact.","tokens_in":29682,"feed_emoji":"🧠","tokens_out":8098,"duration_ms":75233,"temperature":0.7,"pith_summary":"This paper claims that a strictly local, excitatory Hebbian learning rule — the Oja-style update of Eq. (5) with nonnegative weights and competitive normalization — allocates representational resources more efficiently than backpropagation (BP) and difference target propagation (DDTP) when networks are constrained by sparsity and Dale's law. Efficiency is measured by Task-Information Cost, CTI = I(Z;H)/I(Z;Y), the amount of input information retained in a VIB latent code per unit of task-relevant information, estimated post hoc on frozen representations. On three audiovisual benchmarks (AVE, Kinetics-Sounds, VGGSound100) with fixed upstream embeddings, Hebbian-trained MLPs reach substantially lower CTI than sparse BP and DDTP (geometric ratios 5.4 and 6.7) at comparable or slightly lower Top-1 accuracy, and their representations degrade less under magnitude pruning. If the metric is trustworthy, the result supports viewing Hebbian plasticity as a local mechanism for synaptic resource allocation under metabolic constraints, not as a general accuracy-maximizing strategy; the authors themselves stress CTI is a comparative proxy, not a direct energy measure, and limit the claim to shallow and five-layer compressed settings.","feed_headline":"Local Hebbian rule cuts information cost of neural codes","feed_subtitle":"Hebbian representations keep less input information per unit of task-relevant information than sparse backprop or DDTP.","key_machinery":"The carrying mechanism is the constrained Hebbian update of Eq. (5), an Oja-style neural-PCA rule whose subtractive term −η z_j Σ_k z_k w_ik decorrelates postsynaptic activity. Applied with nonnegative weights, min–max rescaling of activations to [0,1], and per-layer z-score normalization, it yields sparse, Dale's-law-compliant, decorrelated hidden codes. The post hoc Variational Information Bottleneck (VIB) module — a diagonal-Gaussian stochastic encoder trained with a KL-to-prior compression penalty and a linear decoder — converts each frozen representation into a latent Z; CTI = I(Z;H)/I(Z;Y), using the KL proxy for I(Z;H) and a variational lower bound for I(Z;Y), turns that code into a s","core_discovery":"Under matched sparsity (~10% connectivity), nonnegativity, and identical MLP capacity, a local excitatory Hebbian update yields post hoc VIB codes with Task-Information Cost (CTI) roughly 9–23 across datasets and architectures, versus 25–344 for sparse BP and 46–210 for DDTP: as much task-relevant information, far less retained input information. Paired log-ratio tests are significant after Holm correction; accuracy is close but not uniformly higher. The authors frame this as a cost-performance trade-off, note that shallow nonnegative BP matches the low-CTI regime but fails deep, and report a ten-hidden-layer Hebbian control that does not stabilize. CTI is explicitly a comparative proxy, not","pith_inferences":["The Gaussian VIB posterior may fit the Hebbian activations' zero-centered, bounded [0,1] shape better than it fits BP/DDTP activations, so the CTI gap could shrink or vanish under a nonparametric estimator of I(Z;H); checking this is the most direct test of the resource-allocation interpretation.","If CTI tracks synaptic maintenance cost as the paper's thermodynamic analogy suggests, Hebbian-trained networks should show measurably lower energy use than BP/DDTP at matched sparsity on neuromorphic hardware; this is a concrete hardware prediction the paper does not test.","The ten-hidden-layer failure suggests the bottleneck readout, not the Hebbian rule, may be removing task-relevant information at depth; adding an explicit inhibitory population (≈20%) or feedback connections, as the paper sketches, could recover depth while preserving low CTI.","Because the inputs are fixed MAE embeddings, the result is about downstream associative plasticity, not end-to-end feature discovery; applying the same assay to raw audiovisual streams would test whether the resource-allocation advantage survives upstream learning."],"forward_implications":["A purely local Hebbian rule operating under cortical-like sparsity can keep task-relevant information without global error signals, which would mean brains can assemble efficient associative codes without backpropagation.","Representational cost can be compared across learning rules independently of raw accuracy: two models with the same Top-1 accuracy can differ by 4–5× in CTI, so accuracy alone is insufficient to evaluate biologically constrained learners.","Hebbian-trained representations tolerate post hoc pruning down to ~10% connectivity with little accuracy loss, while BP and DDTP degrade earlier, implying the learned connectivity is already close to its functional support.","Bimodal audiovisual inputs can raise accuracy without proportionally raising CTI, suggesting that cross-modal integration can be representationally inexpensive under this proxy."],"fun_headline_variants":["Hebbian learning cuts info cost of neural codes","Hebbian rule: same task info, less input info retained","Hebbian plasticity: more efficient neural coding than sparse backprop","Hebbian learning shifts cost-performance trade-off for neural codes","Hebbian codes hold task info with less input data"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The load-bearing premise, spelled out in Secs. 3.2 and 3.6.1, is that CTI = I(Z;H)/I(Z;Y) — an upper-bound KL proxy divided by a variational lower bound — is a valid and comparable measure of representational cost across learning rules; if the Hebbian activations' zero-centered bounded shape makes the Gaussian VIB fit artificially better, the reported CTI advantage is a measurement artifact rather than a property of the rule.","fun_headline_variants_meta":{"raw":{"variants":["Hebbian learning cuts info cost of neural codes","Hebbian rule: same task info, less input info retained","Hebbian plasticity: more efficient neural coding than sparse backprop","Hebbian learning shifts cost-performance trade-off for neural codes","Hebbian codes hold task info with less input data"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001693,"raw_usage":{"total_tokens":6587,"prompt_tokens":827,"completion_tokens":5760,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":571,"completion_tokens_details":{"reasoning_tokens":5690}},"tokens_in":571,"tokens_out":5760,"duration_ms":40335,"temperature":1.0,"reasoning_tokens":5690,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-01T21:33:19.219060+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take the three frozen representation sets (Hebbian, sparse BP, DDTP) for one dataset and architecture, and compute I(Z;H) not with the KL-to-N(0,I) proxy but with the aggregated-posterior KL or a nonparametric estimator (e.g., k-NN) on samples from the VIB encoder, then recompute CTI at matched β and accuracy. If the Hebbian CTI advantage over BP/DDTP vanishes or reverses, the paper's resource-allocation conclusion is an estimation artifact.","supporting_citations":[],"review_version":1}