{"id":"9df038bf-ad8d-4f02-a72d-095c9b4cb128","arxiv_id":"1908.09775","paper_version":1,"verdict":"REJECT","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"high","formal_verification":"none","parameter_count":1,"one_line_summary":"A multi-path wavelet neural network with two learnable filter angles per wavelet neuron is claimed to reach 94.87% on CIFAR-10 with 264k parameters and no augmentation, but the supporting evidence is inconsistent and incomplete.","lead":"The paper proposes an image classifier made of several parallel wavelet decomposition chains whose filter shapes are tuned by backpropagation, followed by two small fully connected layers. It reports high accuracy with only 264,000 parameters on MNIST, CIFAR-10, CIFAR-100, and SVHN, but the experimental claims are not backed by released code or a complete derivation.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The central reported result is internally inconsistent: Table 1 gives 0.89% for the selected 264K model on MNIST, Table 2 gives 0.27% with no explanation; the outperformance claim is therefore unsupported until this is resolved.","rationale":"The paper's aim is to show that a learnable wavelet network with far fewer parameters can match or beat deep learning baselines. The central claim depends on the accuracy and parameter counts reported in Tables 2-4. I looked first for the weakest link in that chain and found an internal contradiction: the architecture selected in Table 1, the 8-path 264K model, has 0.89% MNIST error, but the same model in Table 2 has 0.27%. The text offers no explanation for this discrepancy, so the headline result is not internally reproducible even in principle. This is more fundamental than the reader's identified weakest assumption about baseline protocol, because even if all baselines were perfectly matched, the paper's own selected model would still have an unexplained three-fold difference in its primary result. The parameter-count issue reinforces this: the described architecture with 8 paths and two fully connected layers appears to require more than 264K trainable parameters once the flattened DWT outputs are connected, so the claimed parameter efficiency is also not established. I do not see a way to accept the central outperformance claim while this inconsistency stands. I therefore keep the reader's REJECT verdict, though my route to it is through an internal inconsistency rather than the baseline-matching assumption. A concrete reproduction of the selected configuration on MNIST would settle whether the reported 0.27% is a typo, a different model, or a real result; without that, the paper cannot support its conclusion.","tokens_in":10126,"tokens_out":6645,"duration_ms":69491,"concrete_test":"Independently re-run the selected 8-path configuration on MNIST under the protocol described in Section 4.1 (Adam, learning rate 0.01 with exponential decay, dropout keep probability 0.8, no data augmentation, 10 repeats), and record (a) the final test error and (b) the actual number of trainable parameters from the TensorFlow model. If the error is approximately 0.89% as in Table 1 rather than 0.27% as in Table 2, or if the parameter count is not 264K, the headline performance claim is contradicted by the paper's own architecture-selection data. If the original code is unavailable, a minimal reimplementation from the text is sufficient because the architecture is fully specified.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The most load-bearing concern is not the baseline comparison but the internal inconsistency in the model's own reported results. Section 4.1 says the architecture with 8 parallel paths was selected 'since it gave the best performance' based on Table 1. Table 1 reports that configuration (8 paths, 264K parameters, two fully connected layers) with 0.89% error on MNIST. Yet Table 2, for the identical 264K Learnable Wavelet Neural Network on the same MNIST task, reports 0.27% error (the text even says 0.29%). No additional training procedure, hyperparameter change, ensembling, or data augmentation is described between these two tables. Since every comparative claim in the paper, including the conclusion that the model 'outperforms all the compared deep learning models while training only 264K parameters,' depends on this single number, an unexplained three-fold improvement between the architecture-selection table and the headline table makes the central claim unsupported as written. The parameter count is also never reconciled with the architecture description: with 8 paths each producing three levels of DWT feature maps and then two fully connected layers of 32 units, the flattened input alone appears to exceed the reported 264K budget. Both issues are internal and independent of whether baseline numbers are taken at face value.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a multi-path wavelet neural network with learnable filter parameters for image classification. Each path consists of three levels of 2D discrete wavelet transform using length-6 filters parameterized by two angles α and β, followed by two fully connected layers. Several configurations are evaluated on MNIST; the 8-path, 264K-parameter variant is selected and tested on CIFAR-10, CIFAR-100, and SVHN. The authors report near state-of-the-art accuracy with far fewer parameters than conventional deep models and claim to outperform the compared baselines without data augmentation (except SVHN).","tokens_in":10420,"tokens_out":10653,"duration_ms":100737,"significance":"If the results were reproducible and the parameter counts accurate, this would be a valuable exploration of learnable wavelets as a parameter-efficient alternative to convolutional filters. The idea is interesting and the parameter-savings claim is appealing. However, as written, the manuscript contains internal inconsistencies in its own headline numbers, unverifiable baseline accuracies, and no derivation of the learning rule for the wavelet parameters, so the contribution cannot be assessed from the submitted text.","major_comments":[{"comment":"The same 8-path, 264K-parameter architecture is reported with a 0.89% error rate on MNIST in Table 1 but a 0.27% error rate (0.29% in the text) in Table 2, with no change in training procedure described. This unexplained three-fold performance gap makes the headline result unsupported.","section":"Section 4.1, Tables 1 and 2"},{"comment":"The reported parameter count of 264K is inconsistent with the architecture description. With 8 paths, each containing 3 wavelet neurons, an MNIST input of 28x28x1 produces either 4608 features (if only the final decomposition level is used) or 17,152 features (if all three levels are concatenated). Two fully connected layers of 32 units then require approximately 148K or 549K parameters, respectively, not 264K; for CIFAR inputs of 32x32x3 the required count is even larger. The paper does not clarify which concatenation scheme is used or how 264K is obtained.","section":"Section 3.4 and Table 1"},{"comment":"Several baseline accuracies do not match the cited sources. For example, AlexNet [19] is credited with 94.19% on CIFAR-10, but the cited paper reports ImageNet results, not CIFAR-10, and typical CIFAR-10 AlexNet implementations with data augmentation attain lower accuracy. Similarly, FitNet [45] is credited with 94.61%, whereas the original FitNets paper reports about 91.6% on CIFAR-10. Without consistent and verifiable training protocols for the baselines, the claimed superiority over them is not established.","section":"Section 4.2, Table 3"},{"comment":"The paper states that α and β are updated with backpropagation but provides no derivation of the gradients through the wavelet decomposition, including the downsampling steps and the nonlinear dependence of h(n) on α and β in Eq. (17). A concrete gradient computation is necessary to support the central claim that the wavelet filter angles are learnable.","section":"Section 3.3"},{"comment":"The conclusion asserts that the network 'has been able to outperform all the compared deep learning models,' but Table 2 shows that DropConnect [38] achieves a lower MNIST error rate (0.21% vs the reported 0.27%/0.29%), and the text in Section 4.2 acknowledges that the MNIST result is second best. The conclusion is thus internally contradictory.","section":"Section 5 and Table 2"},{"comment":"The orthonormality claim for the length-6 coefficients is not established. Eq. (16) lists only three conditions, but a quadrature mirror filter of length 6 must also satisfy the shift-orthogonality condition for k=2, e.g., h(0)h(4)+h(1)h(5)=0. The paper neither verifies that Eq. (17) satisfies this condition nor states that it is not required, leaving the theoretical basis of the wavelet system incomplete.","section":"Section 3.3, Eqs. (16) and (17)"}],"minor_comments":[{"comment":"The phrase 'preprocesses the data in th wavelet domain' contains a typo; it should be 'in the wavelet domain.'","section":"Section 2, paragraph 5"},{"comment":"The VGG16 CIFAR-10 entry is written as '92.45' without a percent sign, and the table's formatting is inconsistent with the other entries.","section":"Table 3"},{"comment":"The training description lacks the number of epochs, batch size, and the exact learning-rate decay schedule, which prevents replication.","section":"Section 4.1"},{"comment":"The SVHN dataset is cited as [43] in Section 4.1, but reference [43] is the DenseNet paper; SVHN is described in reference [41].","section":"References and Section 4.1"},{"comment":"The method of concatenating feature maps from the wavelet paths is not specified precisely, in particular whether all decomposition levels are concatenated or only the final level; this ambiguity contributes to the parameter-count problem.","section":"Section 3.4"}],"recommendation":"reject","confidential_remarks":"This manuscript appears to be a preliminary preprint rather than a finished journal article. The internal inconsistency between Tables 1 and 2 is sufficient to reject the central claim, and several baseline numbers are not traceable to the cited papers. The authors should be encouraged to share code, exact training protocols, and corrected parameter calculations before resubmission."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"I read the multi-path wavelet network paper. The idea has a good core: replace fixed wavelet filters with two learnable angles (α, β), stack several wavelet decomposition paths in parallel, and feed the concatenated coefficients into two fully connected layers. The paper positions itself honestly in the wavelet-neural-network literature and acknowledges that each ingredient—learnable wavelets, wavelet-domain layers, multi-path features—exists in earlier work. It even notes its own limitations (sigmoid distortions, computational cost). That is more than many submissions do.\n\nThe problem is the numbers. The architecture search in Table 1 selects the 8-path, 264K-parameter network because it gets 0.89% error on MNIST. Table 2 reports 0.27% for the identical network on the same task (the text says 0.29%). No change in training, architecture, or augmentation is described in between. That internal inconsistency alone makes the headline claim unsupported. The stress-test note is correct: this is the load-bearing issue.\n\nThe parameter count is also suspect. Eight paths, each with three levels of DWT on a 28×28 input, followed by two dense layers of 32 units—a rough count gives around 150K parameters, not 264K. The paper never shows its arithmetic. The baseline comparison numbers don't match the cited sources either; for example, AlexNet at 94.19% on CIFAR-10 without augmentation is not what the original paper reports. Since the whole \"outperforms all\" claim rests on those comparisons, this is a second load-bearing flaw.\n\nFinally, the backpropagation update for α and β is stated but never derived. That would be a minor gap if code were provided, but there is none. No error bars either, just averages over ten runs.\n\nSo the paper is a reasonable sketch of a parameter-efficient wavelet architecture, but the evidence as written does not support the conclusions. The internal MNIST inconsistency and the unverified baselines are not minor polishing issues; they invalidate the central claim. I would not bring this to a reading group, and I would not cite it. If I were the editor, I'd desk reject until the authors reconcile their numbers and provide a reproducible derivation or code.","headline":"The paper's learnable-wavelet idea has a fine core, but the reported results don't survive internal consistency checks: the same 8-path, 264K-parameter MNIST configuration gets 0.89% error in Table 1 and 0.27% in Table 2, and the parameter count doesn't add up.","tokens_in":10914,"tokens_out":3740,"would_cite":false,"duration_ms":37319,"reading_group":"no","serious_thinker":"no","would_accept_peer_review":false},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"A multi-path wavelet neural network with learnable wavelet parameters classifies images at competitive accuracy using only 264,000 trainable parameters and no data augmentation.","keywords":["artificial neural networks","wavelet transform","parameterization","image classification","multi-path architecture","learnable wavelets","low-parameter models","discrete wavelet transform"],"falsifier":"Retrain AlexNet, VGG16, ResNet-110, and DenseNet on CIFAR-10 with the same no-augmentation pipeline used here and compare against the table's 94.19%, 92.45%, 91.57%, and 93.95% figures. If the baselines differ materially, the 'outperforms all' conclusion does not survive. Independently, re-running the eight-path network on CIFAR-10 and checking whether 94.87% accuracy reproduces would test the method itself.","tokens_in":9948,"feed_emoji":"🌊","tokens_out":10243,"duration_ms":93430,"temperature":0.7,"pith_summary":"The paper tries to establish that a neural network whose feature extractor is a learnable discrete wavelet transform can classify images as accurately as much larger deep models while training orders of magnitude fewer parameters. The authors parameterize the wavelet scaling filter by two angles, $\\alpha$ and $\\beta$, make those angles trainable through backpropagation, and arrange wavelet neurons in parallel decomposition paths followed by two fully connected layers. On MNIST, CIFAR-10, CIFAR-100, and SVHN, they report that an 8-path version with 264K total parameters outperforms every compared baseline except DropConnect on MNIST, with no data augmentation on the first three datasets. This matters because parameter count is a major driver of deep learning's data and compute costs, and the result suggests structured transforms can carry much of the feature-extraction burden.","feed_headline":"264K-parameter wavelet net tops deep baselines","feed_subtitle":"Two learnable angles per wavelet filter let a tiny network outscore million-parameter models on MNIST, CIFAR, and SVHN.","key_machinery":"The load-bearing object is the parameterized length-6 scaling filter. The coefficients $h(n)$ are constrained to satisfy the quadrature-mirror conditions—sum $\\sqrt{2}$, unit norm, and orthogonality—and are rewritten in terms of two learnable angles $\\alpha$ and $\\beta$, so a small, differentiable set of parameters determines the whole wavelet system. A wavelet neuron applies the corresponding 2D discrete wavelet transform to an $n\\times n\\times d$ input and produces an $n/2\\times n/2\\times 4d$ output; stacking three neurons gives one path, and the winning architecture uses eight parallel paths. The two angles per neuron are the only trainable parameters in the decomposition stage, with all other parameters residing in the two fully connected layers.","core_discovery":"The central claim is that learnable wavelet decompositions can replace convolutional feature extraction without sacrificing classification accuracy. A length-6 scaling filter sequence $h(0),\\ldots,h(5)$ is expressed through two angular parameters $\\alpha$ and $\\beta$, and each wavelet neuron applies the resulting 2D discrete wavelet transform to an input, outputting one approximation and three detail coefficient maps at half resolution. Three such neurons stacked form a path, and the selected architecture runs eight paths in parallel, concatenates the final coefficient maps, and passes them to two fully connected layers of 32 units each. The authors report 0.27% error on MNIST, 94.87% accuracy on CIFAR-10, 81.22% on CIFAR-100, and 97.51% on SVHN with 264K parameters, which they interpret as evidence that drastically smaller networks can remain competitive with influential deep models.","pith_inferences":["An ablation that freezes $\\alpha$ and $\\beta$ and trains only the fully connected layers would quantify how much of the accuracy actually comes from learning the wavelet parameters.","The same differentiable parameterization could be applied to regression or segmentation tasks, since the transform is not classification-specific.","Because the model trains on a Jetson TX2, the architecture is in principle edge-deployable; benchmarking inference speed and memory would test that promise directly.","One could initialize different paths with different $\\alpha$ and $\\beta$ values and inspect whether they specialize to distinct frequency bands, connecting the approach to multi-scale analysis."],"forward_implications":["If the result reproduces, a convolutional-free feature extractor built from learnable wavelets can reach competitive accuracy on small and medium image benchmarks with far fewer trainable parameters.","Because each wavelet neuron adds only two learnable parameters, widening a network with more parallel paths increases capacity without proportionally increasing the parameter count.","The two-angle parameterization gives a recipe for making other fixed filterbanks learnable, as long as the coefficient sequence satisfies the stated conditions.","The reported 264K-parameter model would be inexpensive to deploy on memory-limited hardware, a direct corollary of the parameter counts the paper reports."],"supporting_citations":[{"why":"Supplies the necessary conditions on the scaling filter coefficients that the two-angle parameterization must satisfy.","marker":"[5]"},{"why":"Backpropagation algorithm used to update the learnable wavelet parameters during training.","marker":"[30]"},{"why":"AlexNet baseline in the comparison tables, with 61M parameters.","marker":"[19]"},{"why":"VGG16 baseline in the comparison tables, with 138M parameters.","marker":"[33]"},{"why":"ResNet baseline in the comparison tables, representing residual deep networks.","marker":"[15]"},{"why":"Inception V3 baseline in the comparison tables.","marker":"[37]"},{"why":"DropConnect baseline that defines the state-of-the-art error rate on MNIST, the one baseline not beaten.","marker":"[38]"},{"why":"DenseNet baseline used in the CIFAR-10, CIFAR-100, and SVHN comparisons.","marker":"[43]"},{"why":"MNIST dataset used for the preliminary architecture search and the MNIST comparison.","marker":"[20]"},{"why":"CIFAR-10 and CIFAR-100 datasets used for the main no-augmentation comparisons.","marker":"[42]"}],"fun_headline_variants":["264K-param wavelet net matches million-param accuracy","Learnable wavelet filters cut net to 264K, keep accuracy","Multi-path wavelet net: 264K params, strong CIFAR and SVHN","Two-angle wavelets build 264K net rivaling deep models"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The comparison with baseline deep models assumes that the quoted baseline accuracies were obtained under the same training protocol as the proposed network, especially no data augmentation on MNIST, CIFAR-10, and CIFAR-100; if that assumption fails, the reported advantage may be an artifact of comparison rather than of architecture.","fun_headline_variants_meta":{"raw":{"variants":["264K-param wavelet net matches million-param accuracy","Learnable wavelet filters cut net to 264K, keep accuracy","Multi-path wavelet net: 264K params, strong CIFAR and SVHN","Two-angle wavelets build 264K net rivaling deep models"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000241,"raw_usage":{"total_tokens":1479,"prompt_tokens":857,"completion_tokens":622,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":473,"completion_tokens_details":{"reasoning_tokens":545}},"tokens_in":473,"tokens_out":622,"duration_ms":6999,"temperature":1.0,"reasoning_tokens":545,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T11:01:45.314825+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Retrain AlexNet, VGG16, ResNet-110, and DenseNet on CIFAR-10 with the same no-augmentation pipeline used here and compare against the table's 94.19%, 92.45%, 91.57%, and 93.95% figures. If the baselines differ materially, the 'outperforms all' conclusion does not survive. Independently, re-running the eight-path network on CIFAR-10 and checking whether 94.87% accuracy reproduces would test the method itself.","supporting_citations":[{"cited_title":"S Burrus, Ramesh A Gopinath, and Haitao Guo","cited_arxiv_id":null,"evidence_quote":"Supplies the necessary conditions on the scaling filter coefficients that the two-angle parameterization must satisfy."},{"cited_title":"Learning repre- sentations by back-propagating errors","cited_arxiv_id":null,"evidence_quote":"Backpropagation algorithm used to update the learnable wavelet parameters during training."},{"cited_title":"Imagenet classification with deep convolutional neural networks","cited_arxiv_id":null,"evidence_quote":"AlexNet baseline in the comparison tables, with 61M parameters."},{"cited_title":"Regularization of neural networks using dropconnect","cited_arxiv_id":null,"evidence_quote":"DropConnect baseline that defines the state-of-the-art error rate on MNIST, the one baseline not beaten."},{"cited_title":"Learning multiple layers of features from tiny images, 2009","cited_arxiv_id":null,"evidence_quote":"CIFAR-10 and CIFAR-100 datasets used for the main no-augmentation comparisons."}],"review_version":1}