{"id":"9c7bf9ae-6819-407d-a62b-455d55989190","arxiv_id":"1908.08987","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"A progressively trained classifier GAN classifies noisy handwritten digits and Bangla characters directly, beating previous baselines on several noisy benchmarks.","lead":"Combining two established AI techniques, progressive GAN training and auxiliary classifier GANs, this paper builds a single model that classifies noisy handwritten characters without first denoising the image. The authors report accuracy gains of one to six percentage points over earlier methods on three noisy handwriting benchmarks.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Reported SOTA margins on Bangla datasets may be confounded by input resolution: PCGAN downsamples noisy 32×32 images to 28×28 (§4.2), attenuating AWGN/Motion noise, while cited baselines were evaluated on 32×32; no ablation isolates progressive training.","rationale":"The paper's contribution is an empirical method and a set of headline numbers. For the central claim to hold, two conditions must be true: the reported accuracies must reflect a fair comparison against prior work, and the progressive-training component must be responsible for the robustness the authors describe. Neither condition is secured in the manuscript. The input-resolution mismatch is the most load-bearing weakness because it does not merely under-determine the mechanism; it raises the possibility that the headline superiority itself is an artifact of preprocessing the authors claim not to need. Downsampling 32×32 noisy images to 28×28 is a mild low-pass filter, and the pattern of margins across noise types matches what that filter would predict. This is a hypothesis, not a demonstrated failure, which is exactly why the missing matched-resolution re-run matters; the paper gives no way to test it. The second weakness, the absence of a fixed-resolution ACGAN ablation, is the one the reader identified; I agree it is real, and it compounds the first. I also note the positive evidence: results are consistent across all nine dataset/noise combinations, McNemar tests are significant for the combinations reported, the architecture is described in enough detail to attempt replication, and the instability on Contrast noise is acknowledged. Offsetting facts: no code, seeds, or error bars; the GAN objective in Equation (1) is garbled; and the progressive scheme omits the fade-in procedure of Karras et al. [14], which was needed in the original method to avoid abrupt resolution jumps. None of these is needed to reach the verdict; the comparison confound and missing ablation are sufficient. I recommend UNVERDICTED rather than CONDITIONAL because the conditions that would rescue the paper (matched-resolution baselines and an ablation) are exactly the evidence required to know whether the headline claim is true at all, and the current text neither provides it nor flags the confound. This is a stronger position than 'acceptable if code is released': the reported numbers themselves are in question, and the manuscript does not permit a reader to resolve that.","tokens_in":9295,"tokens_out":16207,"duration_ms":169185,"concrete_test":"Re-benchmark at matched input conditioning and run the missing ablation. (1) Downsample the 32×32 Noisy Bangla test sets to 28×28 with the exact interpolation PCGAN uses, then evaluate the Karki et al. [13] saliency network and a plain CNN on these downsampled inputs; if either reaches 79.85% (AWGN) or 89.54% (Motion), the reported margins come from the downsample, not the method. (2) Train an ACGAN with the identical discriminator architecture directly at 28×28, skipping the 7×7 and 14×14 modules while keeping optimizer, epochs, and data fixed; if it matches PCGAN, progressive training is not the active ingredient claimed in Section 3.2. Both runs require releasing seeds, code, and per-run spreads, since the manuscript provides none.","verdict_should_be":"UNVERDICTED","load_bearing_attack":"The paper's central empirical claim — state-of-the-art noisy-character accuracy without denoising — rests on a comparison that is not input-matched. Section 4.1 states that the Bangla Numeral and Character datasets are 32×32; Section 4.2 states that these images are resized to 28×28 before being input to PCGAN-CHAR. The baselines cited in Tables 1–2 (Basu et al. [3]; Dropconnect and Karki et al. [13]) report accuracies obtained on the original 32×32 noisy images. Downsampling is a low-pass operation: averaging attenuates independent noise substantially (per-pixel variance drops up to ~4x for a 2×2 box average and still markedly for bilinear interpolation), so PCGAN effectively classifies less noisy inputs than the baselines it is compared with. The margin pattern is consistent with this confound: the largest claimed gains on Bangla Characters are on Motion (+5.95%) and AWGN (+3.11%), the noise types that averaging suppresses, while the only loss is on Contrast (−1.25%), whose multiplicative component survives downsampling; on MNIST, where no resizing occurs, the gains are smaller (0.62–2.21%). The paper never states the baseline input resolutions or controls for this. Compounding this, the mechanism claim of Section 3.2 — that resolution-by-resolution training yields noise robustness — is untested: there is no ablation of the same ACGAN discriminator trained directly at 28×28, so the gains cannot be attributed to progressive training rather than the auxiliary classifier, exposure to generator outputs, dropout/batch norm, or training length. Finally, the McNemar tests in Section 4.3 omit Motion for every dataset and omit Contrast for Bangla Characters, leaving the single largest claimed margin (Motion, +5.95%) without a reported significance test.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes PCGAN-CHAR, a progressively trained auxiliary classifier GAN for classifying noisy handwritten Bangla characters and digits. The discriminator is trained at increasing resolutions (7×7, 14×14, 28×28) and then used as a classifier, with the claim that low-resolution generic features are noise-resistant and that the method can classify raw noisy data without a separate denoising step. The authors report accuracy on three datasets (Noisy Bangla Numeral, Noisy Bangla Characters, Noisy MNIST) under three noise types (AWGN, Contrast, Motion), compare against prior published methods, and provide McNemar tests for several conditions. The headline results are state-of-the-art accuracy on most conditions, with the largest gains on the Bangla datasets.","tokens_in":9635,"tokens_out":3458,"duration_ms":39786,"significance":"If the claims are supported, the work has practical value: an all-in-one classifier that avoids explicit denoising could simplify noisy-character pipelines, and progressive training as a robustness mechanism is an interesting idea. The paper is also commendable for evaluating on multiple datasets and noise types and for including statistical significance tests on several comparisons. However, the significance currently rests on two unverified pillars: (i) that the reported gains come from progressive training rather than from the ACGAN architecture or from input preprocessing, and (ii) that the comparisons to prior work are input-matched. The paper does not provide the ablation or controlled comparison needed to establish these points, so the empirical contribution is not yet convincingly separated from known confounds. The strengths are the breadth of experiments, the direct classification setup, and the presence of significance tests; the weaknesses are the missing ablation and the unresolved resolution mismatch.","major_comments":[{"comment":"The comparison is not input-matched. Section 4.1 states that the Bangla Numeral and Character images are 32×32, while Section 4.2 states that they are resized to 28×28 before being fed to PCGAN-CHAR. Downsampling is a low-pass operation and attenuates independent additive noise, so PCGAN-CHAR is effectively evaluated on less noisy inputs than the baselines if the baselines used the original 32×32 images, as the cited works appear to have done. The paper does not state the input resolution used for Basu et al. [3], Dropconnect [13], or Karki et al. [13], nor does it report results for those baselines on resized inputs. This confound is especially relevant to the largest claimed margins (AWGN +3.11%, Motion +5.95% on Bangla Characters) and to the only loss (Contrast −1.25%), whose multiplicative component would survive downsampling differently. Please add a controlled comparison: evaluate the baselines on 28×28 inputs, or evaluate PCGAN-CHAR on 32×32 inputs, or both, and report the resulting accuracies.","section":"§4.1–4.2, Tables 1–2"},{"comment":"There is no ablation that isolates progressive training. The paper's central mechanism claim is that resolution-by-resolution training makes the discriminator noise-robust (Section 3.2), but the reported experiments only compare against prior published methods. The natural control is the same ACGAN discriminator trained directly at 28×28 without progressive stages, with all other components (architecture, data, loss weights, training budget) held fixed. Without this ablation, the gains cannot be attributed to progressive training rather than to the auxiliary classifier, data augmentation implicit in the GAN objective, or longer training. Please add this ablation, and ideally also a version without the auxiliary classifier, to support the stated mechanism.","section":"§3.2, Algorithm 1, Tables 1–3"},{"comment":"The claim that the framework 'can directly classify raw noisy data without any preprocessing' is inaccurate as stated. Resizing 32×32 images to 28×28, as described in Section 4.2, is itself a preprocessing operation, and for the noise types considered here it tends to suppress noise. This is not 'denoising' in the sense of a learned reconstruction, but it is a resolution change that can affect accuracy. Please either train and evaluate at the native 32×32 resolution, or characterize the effect of resizing on the reported accuracies, or revise the claim to acknowledge this preprocessing step.","section":"§1 and §4.2"},{"comment":"The empirical evaluation omits information needed to assess reliability and reproducibility. The paper does not report learning rate, batch size, optimizer, number of epochs per module, dropout rate, filter counts, or the relative weights of Ldiscern and Lclass in Algorithm 1. It also reports a single run per condition with no error bars. This is a particular concern for claimed gains as small as 0.62% (Noisy MNIST Motion), where run-to-run variation could change the conclusion. Please provide the full hyperparameter settings, the training schedule, and repeated-run statistics (mean and standard deviation, or confidence intervals) for at least the main comparisons.","section":"§4.3 and Implementation details"}],"minor_comments":[{"comment":"The McNemar test results are reported only for a subset of conditions: no test statistic or p-value is given for the Motion-noise conditions on any dataset, and for Bangla Characters only the AWGN condition is tested. Please complete the reporting or explain the omission.","section":"§4.3"},{"comment":"The notation in Eq. (1) is unclear: the term D(x|G(x|z)) appears to mix the generator input and output. It should presumably be D(G(z)), or the conditional dependencies should be defined explicitly.","section":"§3.1, Eq. (1)"},{"comment":"In Algorithm 1, line 5 samples a batch of n random vectors, but line 8 uses G(Bz,Bl) before Bl is sampled on line 7; later, line 9 samples 2n vectors. This is confusing and should be rewritten so that the batching order and the arguments to G are consistent.","section":"Algorithm 1"},{"comment":"There are typographical errors, including 'General Adversarial Networks' in the keywords and abstract, 'disrcriminator' in Section 2, and 'T able 1' in the table heading. A careful proofread is needed.","section":"Throughout"},{"comment":"The architecture diagram does not clearly indicate which layers are shared or how the progressive module weights are transferred. A more detailed description in the caption or a separate table of layer specifications would improve reproducibility.","section":"Fig. 1"}],"recommendation":"major_revision","confidential_remarks":"The paper is an arXiv preprint, and the main baseline [13] is the authors' own earlier work, which is a legitimate comparison. The central issue is not circularity but whether the empirical comparison is fair and whether the claimed mechanism is actually responsible for the gains. I believe the paper is salvageable with additional experiments: an ablation of progressive training, an input-matched baseline comparison, and fuller reporting of hyperparameters and run variability. These are substantial additions, so major revision is appropriate. I would not reject outright, because the idea is plausible and the datasets are standard."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The paper does something reasonable: it takes two existing ideas—ACGAN and progressive GAN training—and points them at noisy handwritten character classification, arguing that the discriminator becomes noise-robust without an explicit denoising step. The experiments cover three datasets and three noise types, and the reported numbers are mostly better than the cited baselines. That is a legitimate contribution, if the numbers hold up.\n\nBut the comparison as presented is not input-matched. The Bangla datasets are 32×32; the paper resizes them to 28×28 before feeding them to PCGAN-CHAR. The baselines in Tables 1–2 are compared at their native 32×32. Downsampling is a low-pass operation, and for AWGN and Motion blur it removes exactly the kind of high-frequency noise that depresses classification accuracy. The result pattern is consistent with this confound: the largest gains are on Motion and AWGN for Bangla Characters (5.95%, 3.11%), the smallest on MNIST, which is already 28×28 (0.62–2.21%). The paper never states baseline input resolutions or controls for this. I would not trust the claimed SOTA margins on the Bangla datasets without a matched-resolution comparison.\n\nThe other soft spot is the mechanism claim. Section 3.2 attributes the noise-robustness to progressive, resolution-by-resolution training. There is no ablation comparing PCGAN-CHAR against the same ACGAN discriminator trained directly at 28×28. For all we know, the gains could come from the auxiliary classifier, exposure to generator outputs, dropout, or longer training. The McNemar tests are also incomplete: no Motion tests are reported for any dataset, and no Contrast test for Bangla Characters—so the largest claimed margin is the one with no significance evidence. No error bars, no code, no hyperparameters.\n\nOn the positive side, the idea of training an all-in-one classifier on noisy data is sensible, the dataset coverage is decent, and the writing is straightforward. The McNemar tests that do appear are appropriate for paired test sets. The paper is not circular or incoherent; it is under-controlled.\n\nThis is worth sending to a serious referee, but the verdict should be conditional on the authors fixing the resolution mismatch, adding a non-progressive ACGAN ablation, and releasing code or at least the exact train/test splits and hyperparameters. As it stands, the paper is a useful cautionary example for empirical comparisons more than a reliable new benchmark.\n\nFor a reading group, it would spark a good discussion about confounds in applied deep learning papers. I would not cite the Bangla accuracy numbers in my own work until the matched-resolution results are out.","headline":"Smart combination of ACGAN and progressive GAN training, but the headline SOTA results are undermined by an input-resolution mismatch and a missing ablation.","tokens_in":10220,"tokens_out":1994,"would_cite":false,"duration_ms":21914,"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":"Progressively training a classifier GAN from low to high resolution lets it classify noisy handwritten characters accurately without any denoising step.","keywords":["generative adversarial networks","progressive growing","noisy handwritten character recognition","Bangla numeral classification","auxiliary classifier GAN","noise robustness","sparse features","MNIST"],"falsifier":"Train the same ACGAN classifier at a fixed 28x28 resolution, with no progressive modules, for the same total number of epochs on Noisy Bangla Characters with AWGN noise. If that fixed-resolution model reaches or exceeds 79.85% accuracy, the progressive-training explanation is not what carries the result.","tokens_in":9085,"feed_emoji":"📝","tokens_out":6611,"duration_ms":63135,"temperature":0.7,"pith_summary":"This paper tries to show that a single classifier can recognize noisy handwritten characters directly from raw pixels, with no denoising step, by training a generative adversarial network progressively from low to high resolution. The claim is that learning the features of each resolution independently makes the model's discriminator robust to noise, because low-resolution features are generic and survive corruption better than fine-grained features. On the three benchmark tasks, the proposed model reports the best published accuracies on eight of nine noise-dataset combinations: for example 96.68% on Noisy Bangla Numeral with added white Gaussian noise, 79.85% on the 50-class Noisy Bangla Characters with the same noise, and 98.43% on Noisy MNIST. On the one exception, Noisy Bangla Characters under contrast noise, it reaches 68.41%, slightly below the prior best of 69.66%. If these results hold, noisy character recognition can skip a separate preprocessing stage entirely.","feed_headline":"Progressive GAN classifier beats prior best on noisy Bangla chars","feed_subtitle":"PCGAN-CHAR skips denoising, trains 7x7 to 28x28, and hits top accuracy on most noisy benchmarks.","key_machinery":"The load-bearing mechanism is the progressively trained ACGAN discriminator, used as the classification network. ACGAN is a GAN variant whose discriminator has an auxiliary head predicting the input's class label in addition to the real/fake decision; progressive growing is the technique of training the generator and discriminator in stages at increasing spatial resolution, here 7x7, 14x14, and 28x28, copying the trained weights into the next stage. What this mechanism does is force the discriminator to learn generic, low-resolution features first and then finer features, so the final classifier has a set of resolution-specialized layers. Since low-resolution features are less disrupted by noise, the resulting classifier can label noisy raw inputs directly. The paper emphasizes that no denoising is performed at any point in training or inference.","core_discovery":"The central discovery is that progressive resolution training turns an Auxiliary Classifier GAN's discriminator into a noise-robust classifier. The model, built from an ACGAN in which the discriminator both distinguishes real from generated images and predicts class labels, is trained in three modules at 7x7, 14x14, and 28x28 resolution, with weights transferred from each module to the next. Because the generator supplies a stream of fake images, the discriminator sees examples ranging from noisy to sharp, and because it learns each resolution in isolation, its lower layers retain generic, noise-resistant features. The reported outcome is best-or-near-best accuracy on noisy Bangla numerals, noisy Bangla characters, and noisy MNIST under white Gaussian, motion-blur, and contrast noise, with statistically significant gains over a pixel-level-reconstruction baseline on the tested settings. The paper interprets these results as showing that an all-in-one classifier can work without explicit or implicit denoising.","pith_inferences":["If progressive training is the active ingredient, then an ablation that trains the same ACGAN at a fixed 28x28 resolution for the same total number of epochs should lose several accuracy points under AWGN and motion noise; the paper does not report such an ablation, so this is a testable prediction rather than a demonstrated result.","The same low-resolution-first idea could be ported to other sparse-feature recognition tasks, such as medical symbols, degraded printed text, or noisy CAPTCHAs, where current pipelines still denoise before classifying.","Because contrast noise is the weakest setting, a natural extension would be to add contrast-degraded real images to the discriminator's training batches at each progressive stage; if contrast robustness is a resolution-specialization issue, this should close the 1.25% gap."],"forward_implications":["Noisy handwritten character recognition can be treated as a single end-to-end classification problem: raw noisy pixels in, class label out, with no denoising preprocessing.","Progressive, resolution-by-resolution training appears to make GAN discriminators more robust to added image noise, a property that could generalize beyond character images.","On 10-class datasets (numerals and MNIST), the method's accuracy stays above 94% for all three noise types, while on the 50-class Bangla character set accuracy drops sharply, showing that class count remains the main difficulty.","The one reported miss, 68.41% versus 69.66% on contrast-degraded Bangla characters, shows the method does not dominate every noise type even when it leads on the other two."],"supporting_citations":[{"why":"supplies the progressive growing technique that trains GAN layers at increasing spatial resolutions.","marker":"[14]"},{"why":"supplies the auxiliary classifier GAN architecture whose discriminator predicts class labels alongside real/fake.","marker":"[20]"},{"why":"provides the noisy Bangla and MNIST datasets, the prior pixel-level reconstruction baseline, and the statistical comparison target.","marker":"[13]"},{"why":"provides the base generative adversarial network formulation and minmax objective.","marker":"[9]"},{"why":"supplies the generic-to-specific transferable feature result that motivates progressive resolution training.","marker":"[23]"},{"why":"is the source of the original Bangla numeral dataset.","marker":"[5]"},{"why":"is the source of the original Bangla character dataset.","marker":"[6]"},{"why":"provides the sparse-feature representation context and a baseline classifier.","marker":"[3]"}],"fun_headline_variants":["Progressive GAN classifier beats prior best on noisy Bangla","No denoising: progressive GAN classifies noisy Bangla chars","Progressive GAN yields noise-robust classifier for Bangla","Train GAN at increasing res to beat noisy Bangla characters","PCGAN-CHAR: progressive GAN classifier for noisy Bangla"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The paper's argument stands on the assumption that progressive resolution-by-resolution training, rather than the ACGAN architecture, the generator's fake-image curriculum, or longer training, is what gives the discriminator its noise robustness; the evaluation compares against prior published methods but does not include an ablation of the model's own components.","fun_headline_variants_meta":{"raw":{"variants":["Progressive GAN classifier beats prior best on noisy Bangla","No denoising: progressive GAN classifies noisy Bangla chars","Progressive GAN yields noise-robust classifier for Bangla","Train GAN at increasing res to beat noisy Bangla characters","PCGAN-CHAR: progressive GAN classifier for noisy Bangla"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00119,"raw_usage":{"total_tokens":4873,"prompt_tokens":870,"completion_tokens":4003,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":486,"completion_tokens_details":{"reasoning_tokens":3913}},"tokens_in":486,"tokens_out":4003,"duration_ms":30657,"temperature":1.0,"reasoning_tokens":3913,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T13:59:02.206407+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Train the same ACGAN classifier at a fixed 28x28 resolution, with no progressive modules, for the same total number of epochs on Noisy Bangla Characters with AWGN noise. If that fixed-resolution model reaches or exceeds 79.85% accuracy, the progressive-training explanation is not what carries the result.","supporting_citations":[{"cited_title":"In: Proceedings of the 34th International Conference on Machine Learning- Volume 70","cited_arxiv_id":null,"evidence_quote":"supplies the auxiliary classifier GAN architecture whose discriminator predicts class labels alongside real/fake."},{"cited_title":"In: Ghahramani, Z., Welling, M., Cortes, C., Lawrence, N.D., Weinberger, K.Q","cited_arxiv_id":null,"evidence_quote":"provides the base generative adversarial network formulation and minmax objective."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"supplies the generic-to-specific transferable feature result that motivates progressive resolution training."},{"cited_title":"IEEE transactions on pat- tern analysis and machine intelligence 31(3), 444–457 (2009) Title Suppressed Due to Excessive Length 13","cited_arxiv_id":null,"evidence_quote":"is the source of the original Bangla numeral dataset."},{"cited_title":"Pattern Analysis and Applications 15(4), 445–458 (2012)","cited_arxiv_id":null,"evidence_quote":"is the source of the original Bangla character dataset."},{"cited_title":"Neural Processing Letters 45(3), 855–867 (2017)","cited_arxiv_id":null,"evidence_quote":"provides the sparse-feature representation context and a baseline classifier."}],"review_version":1}