{"id":"2c157a9e-fb2c-4ad8-83dd-fb274b50f169","arxiv_id":"2501.16677","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":7,"one_line_summary":"A new sparsity loss that binarizes class-specific CNN filter activations during training improves neuro-symbolic rule extraction accuracy by 9% and reduces rule-set size by 53% versus the prior baseline.","lead":"This paper introduces a sparsity loss that pushes a few class-specific filters in a CNN to fully activate and the rest to stay silent during training, so the network's outputs can be rounded to 0/1 for rule extraction with little information loss. The approach improves the accuracy of the resulting interpretable neuro-symbolic model by 9% on average and shrinks the rule set by 53% compared to the previous state of the art, bringing it within 3% of the original CNN's accuracy.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The reported 9%/53% benchmark relies on TS3's random class-filter assignment, yet no variance across random assignments is reported; the headline numbers may not be reproducible.","rationale":"The reader's verdict is CONDITIONAL, and the reader flagged both the untested root-cause claim and the surprising robustness of random filter assignment. I agree that the root-cause attribution is not directly tested, but the more load-bearing threat to the paper's central benchmark is the missing variance around TS3. The paper's headline numbers depend on TS3, and TS3's only substantive difference from TS2 is a randomly initialized P matrix. Reporting an average over 5 runs without stating whether P was resampled, and without error bars, leaves open the possibility that the reported 9% improvement is a lucky draw. This is a concrete, testable concern: recomputing TS3 with multiple random P matrices would either confirm that the assignment is irrelevant (as the paper implies) or reveal that the method is sensitive to a choice the paper treats as arbitrary. I am not recommending rejection because the underlying method is plausible and the comparison to NeSyFOLD-EBP is internally consistent; instead, the paper should be accepted conditionally on supplying this robustness evidence. The proposed test is specific and would settle whether the headline benchmark is stable.","tokens_in":12562,"tokens_out":9647,"duration_ms":106826,"concrete_test":"Run TS3 on P3.1 and GT43 with 10 independent random P matrices, keeping the CNN initialization, training seed, and all hyperparameters fixed so that only the class-filter assignment varies. Report the NeSy accuracy, rule-set size, and fidelity for each run as mean ± standard deviation and worst case. If the standard deviation of accuracy exceeds about 1.5 points, or if the worst-case accuracy is within noise of the NeSyFOLD-EBP baseline (78% on P3.1 and 95% on GT43 in the paper's averages), then the reported benchmark is not robust to the random assignment and the paper must present the full distribution rather than a single averaged point.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central empirical claim is the TS3 result: +9% accuracy and −53% rule-set size versus NeSyFOLD-EBP. TS3 differs from the informed-selection strategy TS2 only in that the P matrix is initialized by randomly assigning K filters to each class. The paper says results are averaged over 5 runs but does not report whether the random P matrix is resampled per run, nor any standard deviation. Since the CNN is initialized from fixed ImageNet weights, the random assignment is the main uncontrolled source of variation. If some assignments yield much worse accuracy, the reported mean overstates the method. Moreover, the paper's explanation for why random beats informed selection is post hoc: it asserts that pretrained weights are not optimal for the sparsity loss, but provides no test of this claim. The success of TS3 could instead come from the sparsity loss acting as a generic feature-separability regularizer, not from the specific class-specific binarization mechanism claimed in Section 1. Without a variance analysis or a comparison to an alternative regularizer, the headline 9% and 53% figures are not yet established as reproducible.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper argues that the accuracy loss incurred by NeSyFOLD-style rule extraction stems from post-training binarization of CNN filter activations. To mitigate this, it proposes a sparsity loss that assigns each class a small set of filters with target activation values of 1 and pushes all other filter outputs toward 0, using a threshold-adjusted sigmoid and a binary cross-entropy term. Five training strategies (TS1--TS5) are evaluated on Places subsets (P2--P10) and GTSRB. The main claim is that TS3, which uses random class--filter assignment, improves NeSy model accuracy by 9% and reduces rule-set size by 53% on average relative to NeSyFOLD-EBP, while coming within 3% of the original CNN's accuracy. The paper also provides qualitative evidence that the learned filters correspond to class-relevant concepts and gives usage guidance for the different strategies.","tokens_in":12858,"tokens_out":4059,"duration_ms":43883,"significance":"If the reported results are robust, the proposed loss is a meaningful step toward making neuro-symbolic image classifiers competitive with black-box CNNs. The method is simple, the experimental setup is clearly described, and the comparison to NeSyFOLD-EBP follows the same pipeline on the same datasets, which is a strength. The claims are also falsifiable through the stated research questions. However, the headline 9% and 53% figures depend on a randomly initialized assignment whose variance is not reported, and the central causal claim that post-training binarization is the root cause of the accuracy gap is not directly tested. These issues currently limit the confidence one can place in the benchmark numbers.","major_comments":[{"comment":"The main accuracy/rule-set-size claim (9% improvement, 53% reduction) uses TS3, where the P matrix is initialized randomly via Method 2 in Section 3.1. The text says results are averaged over 5 runs, but it does not state whether the random assignment is resampled across those runs, and no standard deviation, per-run values, or significance tests are reported for any table. Because the CNN starts from fixed ImageNet weights and the data splits are fixed, the random P matrix is the main uncontrolled source of variation. Without a measure of this variance, the reported means do not establish that the headline numbers are reproducible.","section":"Section 4, Table 1 (TS3)"},{"comment":"The paper states that post-training binarization of filter activations is the root cause of the accuracy loss, but no experiment directly tests this claim. TS2 and TS3 reduce the accuracy gap, but this could also result from the sparsity loss acting as a generic feature-separability regularizer rather than from the specific pseudo-binarization mechanism. An ablation is needed: for example, train with the same sparsity loss but replace the binary target assignment with a non-binarizing feature-separation objective, or compare against a baseline that uses the same thresholds but a different binarization scheme. Without such an ablation, the attribution of the gains to 'binarization during training' is underdetermined.","section":"Section 1 and Section 4 (causal claim)"},{"comment":"The method depends on several hyperparameters -- alpha, beta, h1, h2, K, ratio, and tail -- but no sensitivity analysis is reported. The guidance in Section 6 about when to use TS2 versus TS3 versus TS4 is only meaningful if performance is stable under reasonable variations of these settings. At minimum, the authors should report the effect of beta and h1/h2 on one dataset, and of K on one larger dataset, especially because the threshold subtraction in Equations (2)--(5) is shown to be critical by the degraded TS5 results.","section":"Section 4 (Hyperparameters)"}],"minor_comments":[{"comment":"The caption contains a typo: 'NeSy mdoel' should be 'NeSy model'.","section":"Table 2 caption"},{"comment":"There are small formatting issues: 'FOLD-SE-M[Wang and Gupta, 2024]' and 'FOLD-SE-M [Wang and Gupta, 2024 ]' are inconsistent, and 'abx' should be typeset consistently as a predicate symbol.","section":"Section 2 and 4"},{"comment":"The qualitative filter-analysis result is based on selecting one top filter per class from the rule-set. The paper does not report how many filters were examined or whether the shown examples are representative; a quantitative concept-alignment metric would strengthen this section.","section":"Section 4, Q4"}],"recommendation":"major_revision","confidential_remarks":"The baseline is the authors' own NeSyFOLD-EBP, and no comparison is made to other recent neuro-symbolic rule-extraction methods; an independent baseline or a discussion of why such comparison is not feasible would help the editor assess the novelty claim. Code is not released, which further impedes independent verification of the random-assignment results."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The one thing you should know: the paper introduces a new sparsity loss that trains CNN filters to produce near-binary, class-specific activations, and the experiments show consistent gains over the NeSyFOLD-EBP baseline. That is a real contribution. The second thing: the headline numbers come from TS3, where the class-filter assignment is random, and the paper reports no variance across random assignments. So the 9% accuracy gain and 53% rule-set reduction are plausible but not yet established.\n\nThe method itself is simple and well motivated. Instead of binarizing continuous filter outputs after training, they add a BCE loss that pushes pre-selected filters toward 1 and others toward 0 during training. The five training strategies are a reasonable ablation, and the discussion of why TS3 beats TS2 is honest, even if post hoc. The filter visualization in Section 4 is a nice touch. I also appreciate the scalability experiments on P10 and GT43.\n\nThe soft spots are mostly evidential. First, no error bars or significance tests anywhere. The tables are rounded averages over five runs, and we don't know if the TS3 vs TS2 gap is stable. Second, no code or data release, which makes replication harder. Third, the paper claims the root cause of accuracy loss is post-training binarization, but never tests that by, say, keeping binarization fixed and varying the rule extractor. The success of TS3 with random assignment is also unexplained; the explanation about pretrained weights not being optimal for the sparsity loss is plausible but untested. These are fixable in revision, and none of them kill the core idea.\n\nI disagree with the stress-test note if it implies the random assignment makes the result worthless. The gains are consistent across seven datasets, and TS3's advantage over NE is large. But the note is right that the headline numbers need a variance bound.\n\nWho is this for? Anyone working on interpretable neuro-symbolic classifiers or activation sparsity. It deserves a serious referee: the method is novel enough, the experiments are extensive, and the limitations are addressable. I'd send it out, with a request for error bars, code, and an ablation that directly tests the binarization claim.","headline":"Novel sparsity loss with class-specific binarization targets, but the headline 9%/53% gains lack variance reporting and a direct test of the claimed root cause.","tokens_in":13301,"tokens_out":2324,"would_cite":true,"duration_ms":22984,"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":"A class-specific sparsity loss that trains CNN filters to fire near 0 or 1 reduces neuro-symbolic rule-set size by 53% and lifts accuracy by 9%, closing the gap to the original CNN to about 3%.","keywords":["neuro-symbolic","rule extraction","CNN interpretability","sparsity loss","class-specific filters","binarization","logic programming","image classification"],"falsifier":"Train the same VGG16 with the TS3 sparsity loss, then compare rule extraction from the pseudo-binary sigmoid values versus from the raw continuous L2 norms fed to a rule learner that accepts real-valued inputs (e.g., a soft decision tree). If the continuous-pipeline accuracy is still more than 3% below the CNN, then binarization is not the main source of loss. Alternatively, run TS3 but replace FOLD-SE-M with a standard decision-tree extractor on the same binarized vectors; if the 9% gain over EBP disappears, the benefit is specific to FOLD-SE-M rather than to the binarization.","tokens_in":12400,"feed_emoji":"🧩","tokens_out":9589,"duration_ms":76483,"temperature":0.7,"pith_summary":"The paper claims that the accuracy loss in neuro-symbolic image classification—where a CNN's last-layer filter outputs are converted to 0/1 and a rule-set replaces the fully connected layers—comes mainly from post-training binarization. To fix it, the authors train the CNN with an added sparsity loss that pushes a pre-selected set of class-specific filters toward 1 and all others toward 0, so the outputs are already nearly binary when the rule-set is extracted. On several scene and traffic-sign datasets, the best training strategy raises neuro-symbolic accuracy by 9% on average over the prior state of the art and shrinks the rule-set by 53%, closing the gap to the original CNN to about 3%. This would make interpretable rule-sets a much more viable substitute for black-box CNNs.","feed_headline":"Close the CNN-to-rules gap to 3% with a sparsity loss","feed_subtitle":"By binarizing filters during training, neuro-symbolic models come within 3% of CNN accuracy.","key_machinery":"The key object is the sparsity loss function. For each training image, the L2 norm of each last-layer filter is centered by subtracting a precomputed threshold (a per-filter mean plus scaled standard deviation), passed through a sigmoid, and then compared via binary cross-entropy to a target of 1 for the selected class-specific filters and 0 for all others. The selection is encoded in a filter probability matrix $P$ of shape $C \\times F$, and the thresholds are computed from the training set before the loss is applied. This setup turns filter outputs into pseudo-binary values, so rounding them to 0/1 after training is nearly lossless. The paper tests five training strategies that vary how $P$ is chosen and when the loss is activated; TS3, which uses a random $P$ and applies the loss from the start alongside cross-entropy, gives the best accuracy and rule-set size.","core_discovery":"The paper's central claim is that training a CNN with a class-specific sparsity loss—one that drives a pre-selected set of K filter outputs per class toward 1 and all other filter outputs toward 0 via binary cross-entropy on sigmoid-adjusted L2 norms—makes the filter activations nearly binary before rule extraction. When these rounded activations are fed to the FOLD-SE-M rule learner, the extracted logic program is both more accurate and much smaller than the one produced by the prior EBP-based pipeline. Surprisingly, the best-performing strategy assigns the per-class filter targets randomly rather than by activation strength; the authors interpret this as showing that the binarization pressure itself, not the choice of which filters to keep, does the work. As a result, the neuro-symbolic model comes within 3% of the original CNN's accuracy on average across the tested datasets.","pith_inferences":["We infer that if binarization is indeed the root cause, the sparsity loss should transfer to other rule extractors (e.g., decision trees or rule lists) and other CNN backbones; this is testable and the paper does not run it.","The robustness to random filter assignment hints that the loss's regularizing effect, rather than the semantic content of filters, drives the gains; a useful experiment would compare random maps to adversarially chosen maps to see if any fixed partition works.","The authors' future-work suggestion of extending the loss to Vision Transformers is natural, but the threshold-and-sigmoid construction relies on spatial feature-map norms, so an adaptation would need a different normalization for attention outputs.","In safety-critical settings, the smaller rule-sets combined with answer-set-programming justification could allow formal verification of individual predictions; the paper demonstrates the justification machinery but does not quantify verification cost."],"forward_implications":["If the loss works as described, neuro-symbolic classifiers can match a fine-tuned CNN to within 3% accuracy, making rule-based explanations viable for deployment in high-stakes image classification.","The 53% average reduction in rule-set size means extracted logic programs have far fewer predicates, which lowers the human cost of auditing and verifying the rules.","The finding that a randomly assigned filter-class map (TS3) beats activation-based selection (TS2) suggests the method is robust to filter initialization and may not require dataset-specific hyperparameter tuning for the filter map.","On 10-class and 43-class datasets, the gains over the baseline increase, indicating the sparsity loss scales with label count rather than degrading.","The success of TS4, which uses only the sparsity loss and no cross-entropy, shows that binarized filter features alone can support rule extraction, though with lower fidelity to the original CNN."],"supporting_citations":[{"why":"Defines NeSyFOLD, the baseline framework whose accuracy loss the paper targets, and supplies the rule-extraction pipeline and semantic labelling used in evaluation.","marker":"[Padalkar et al., 2024a]"},{"why":"Provides FOLD-SE-M, the rule-based learning algorithm that converts binarized filter vectors into the logic program rule-set.","marker":"[Wang and Gupta, 2024]"},{"why":"Introduces Elite BackProp (EBP), the class-specific sparse filter training method used by the previous SOTA (NeSyFOLD-EBP) and the direct comparison point for the new sparsity loss.","marker":"[Kasioumis et al., 2021]"},{"why":"Supplies the Places dataset subsets (P2, P3.1, P3.2, P3.3, P5, P10) on which all training strategies are evaluated.","marker":"[Zhou et al., 2017]"},{"why":"Supplies the GTSRB traffic-sign dataset with 43 classes used to test scalability.","marker":"[Stallkamp et al., 2012]"},{"why":"Supplies ImageNet pretraining weights for VGG16, the CNN backbone used across all experiments.","marker":"[Deng et al., 2009]"}],"fun_headline_variants":["Sparsity loss shrinks rule sets 53% and closes accuracy gap to 3%","Train-time binarization beats post-hoc extraction for neuro-symbolic rules","Random filter targets still win: sparsity pressure does the work","Class-specific sparse filters cut rule size by 53% and lift accuracy 9%","Binarize during training: rule sets 53% smaller, within 3% of CNN accuracy"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the accuracy gap between a CNN and its neuro-symbolic surrogate is caused mainly by post-training binarization of filter activations; if the gap comes instead from the rule-learning algorithm or the threshold choices, the sparsity loss would not address the real problem.","fun_headline_variants_meta":{"raw":{"variants":["Sparsity loss shrinks rule sets 53% and closes accuracy gap to 3%","Train-time binarization beats post-hoc extraction for neuro-symbolic rules","Random filter targets still win: sparsity pressure does the work","Class-specific sparse filters cut rule size by 53% and lift accuracy 9%","Binarize during training: rule sets 53% smaller, within 3% of CNN accuracy"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00031,"raw_usage":{"total_tokens":1771,"prompt_tokens":951,"completion_tokens":820,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":567,"completion_tokens_details":{"reasoning_tokens":710}},"tokens_in":567,"tokens_out":820,"duration_ms":7276,"temperature":1.0,"reasoning_tokens":710,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T11:28:16.194733+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Train the same VGG16 with the TS3 sparsity loss, then compare rule extraction from the pseudo-binary sigmoid values versus from the raw continuous L2 norms fed to a rule learner that accepts real-valued inputs (e.g., a soft decision tree). If the continuous-pipeline accuracy is still more than 3% below the CNN, then binarization is not the main source of loss. Alternatively, run TS3 but replace FOLD-SE-M with a standard decision-tree extractor on the same binarized vectors; if the 9% gain over EBP disappears, the benefit is specific to FOLD-SE-M rather than to the binarization.","supporting_citations":[{"cited_title":"Places: A 10 million image database for scene recognition, 2017","cited_arxiv_id":null,"evidence_quote":"Supplies the Places dataset subsets (P2, P3.1, P3.2, P3.3, P5, P10) on which all training strategies are evaluated."},{"cited_title":"Imagenet: A large-scale hierarchical image database","cited_arxiv_id":null,"evidence_quote":"Supplies ImageNet pretraining weights for VGG16, the CNN backbone used across all experiments."}],"review_version":1}