{"id":"4f7b4835-c3fd-401a-b4f0-3898b721ad68","arxiv_id":"2501.15547","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":10,"one_line_summary":"A 14,862-parameter CNN trained with dual original/augmented branches and progressive unfreezing reports 99% on MNIST and 89% on Fashion-MNIST.","lead":"The paper presents a two-stage training recipe in which twin CNNs are trained on original and augmented data, then merged and fine-tuned with progressive unfreezing. On MNIST and Fashion-MNIST it reports 99% and 89% accuracy from a 14,862-parameter model.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The reported parameter counts do not match the final architecture described in §3.5.2: replacing Dense with 1×1 Conv2D gives ~20,622 params for MNIST, while Table 6's 14,862 matches a model that keeps the original Dense layers, so the claimed size and the described method are inconsistent.","rationale":"The reader's weakest assumption (no baselines/ablations) is valid, but I found a more immediate, internally checkable problem. The parameter counts in Table 6 are exactly reproduced only if the original branch Dense(10) layers are kept and their outputs concatenated as 20-vectors; the text says those Dense layers are removed and replaced by 1×1 Conv2D. For MNIST, the 1×1-conv version has 20,622 params, not 14,862; the retained-dense version has 14,862. For CIFAR-10, the same pattern holds (28,022 vs 19,622). Since the paper provides no code, a reader cannot tell which architecture produced the 99%/89%/65% accuracies. This does not prove the results are wrong—the 14,862-parameter retained-dense model is a plausible small CNN—but it makes the central efficiency claim unverifiable as written. I therefore keep the CONDITIONAL verdict: acceptance should require code or a corrected, unambiguous architecture description, plus the baselines/ablation the reader already requested.","tokens_in":11623,"tokens_out":10031,"duration_ms":88366,"concrete_test":"Implement the final MNIST/Fashion model exactly per §3.5.2 in Keras: each branch = Conv2D(10,3,valid)+MaxPool+Conv2D(20,3,valid)+MaxPool, then remove the branch Dense layers and insert Conv2D(10,(1,1),padding='same'); concatenate branch outputs, Flatten, Dense(32), Dense(10), and call model.count_params(). If the result is 20,622 instead of 14,862, the described architecture cannot produce the paper's headline parameter count; request corrected code or a corrected kernel size (e.g., 5×5) and rerun both the parameter count and the test-accuracy evaluation.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The headline efficiency claim ('14,862 parameters, 0.17 MB') cannot be reproduced from the architecture described in §3.5.2. For MNIST/Fashion, after two valid-padding 3×3 convolutions and 2×2 max-pooling the branch feature map is 5×5×20; the original Dense(10) layer therefore has 500×10+10=5,010 params per branch. The text says these dense layers are 'removed and replaced' by 1×1 Conv2D with 10 filters, which would have only 20×10+10=210 params per branch. Then the 1×1-conv outputs (5×5×10 each) concatenate to 5×5×20, flatten to 500, and feed Dense(32) (500×32+32=16,032) plus Dense(10) (330). Total: 2×(100+1,820+210)+16,032+330 = 20,622 params, not 14,862. The reported 14,862 is exactly the count obtained when the original Dense(10) layers are kept and their 10-vectors are concatenated before Dense(32): 2×(100+1,820+5,010)+672+330 = 14,862. The CIFAR-10 count (19,622) matches the retained-dense model as well. Thus either §3.5.2's 1×1 replacement was not actually performed, or Table 6 is for a different model; the paper does not resolve this, and there is no code. The causal attribution to the proposed method is also untested because no baseline or ablation is given, but the parameter-count inconsistency is the more immediate obstacle: as written, the model whose performance is claimed and the model whose size is claimed are not demonstrably the same.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a two-stage training methodology for building small CNNs: first, a dual-input, dual-output model with two identical branches is trained, one branch on the original data and one on an augmented version; second, a single-input model is created by removing/replacing the branch dense layers, concatenating the remaining features, adding a new dense head, and fine-tuning with progressive unfreezing. The manuscript reports 99% test accuracy on MNIST, 89% on fashion MNIST, and 65% on CIFAR-10 with 14,862/19,622 trainable parameters and model sizes of about 0.17–0.21 MB, together with latency and throughput measurements. The abstract and introduction describe these results as state-of-the-art, although Section 4 explicitly states that the experiments are not directly compared with other models.","tokens_in":12077,"tokens_out":3864,"duration_ms":33642,"significance":"If fully supported, the paper would provide a useful data point: a roughly 15k-parameter model that reaches 99% on MNIST and 89% on fashion MNIST is a plausible and potentially useful efficiency result, and the reported test accuracies, classification reports, and confusion matrices are internally consistent. The held-out test-set evaluation is standard, and the paper is commendably transparent about several limitations, including the absence of direct comparisons. However, the central claims are not yet established: the described architecture in Section 3.5.2 does not match the parameter counts in Table 6, there is no baseline or ablation to support the causal attribution to the proposed method, and the 'state-of-the-art' label is unsupported. The contribution is primarily an empirical combination of existing techniques (augmentation, transfer learning, progressive unfreezing) rather than a new conceptual mechanism.","major_comments":[{"comment":"The reported parameter count of 14,862 does not match the architecture described in Section 3.5.2. For a 28×28×1 input, the branch feature map before the final dense layer is 5×5×20; retaining the original Dense(10) layers gives 2×(100+1,820+5,010)+672+330 = 14,862 parameters, whereas replacing those dense layers with 1×1 Conv2D(10) layers as stated gives 2×(100+1,820+210)+16,032+330 = 20,622 parameters. The CIFAR-10 count of 19,622 likewise matches the retained-dense model, not the 1×1-converted model. The paper therefore conflates two different models: the size claimed is for the original architecture, while the described 1×1 conversion produces a larger model. Please correct the text, the table, or provide a model artifact that resolves this discrepancy.","section":"§3.5.2 and Table 6"},{"comment":"The abstract's 'state-of-the-art accuracy' claim is not established by the body of the paper. Section 4 states that the results 'are not directly compared to other models,' and the only comparison offered is a qualitative remark about LeNet. Since 99% on MNIST is a routinely achieved accuracy and many models with comparable or smaller parameter counts exceed it, please replace the 'state-of-the-art' phrasing with a supported claim, or add a comparison table with relevant small CNN baselines.","section":"§4 and Abstract"},{"comment":"No ablation or baseline isolates the contribution of the dual-input design, the concatenation, or progressive unfreezing. For example, a single-branch CNN with the same parameter budget trained on augmented data might reach similar accuracy, which would reduce the methodological contribution to routine augmentation plus a small architecture. Please add at least a single-branch baseline trained on original data and a single-branch baseline trained on augmented data, and, if possible, a version of the proposed pipeline without the final progressive-unfreezing stage.","section":"§3.5 and §4"}],"minor_comments":[{"comment":"The manuscript contains many typographical errors that should be corrected in a copyedit; examples include 'and and filter decomposition' in Section 1, 'T ransfer' and 'modther model' in several places, 'CIF AR-10' throughout, and 'rebuilded-model' in the final paragraph.","section":"Throughout"},{"comment":"The caption calls CIFAR-10 a 'handwritten' dataset; CIFAR-10 is not a handwritten dataset, so the caption should be corrected.","section":"Figure 13 caption"},{"comment":"The text says the model was trained for 20 epochs with early stopping, but the note in the Figure 9 caption says increasing epochs beyond 20 did not help; please clarify the final epoch count and the stopping criterion used for each dataset.","section":"§3.5.3 Stage 1"},{"comment":"Please state how 'Approx. Final Training Accuracy' is computed and whether it is the average over the six folds from Section 3.5.3, since the table currently gives no variance information.","section":"Table 1"}],"recommendation":"major_revision","confidential_remarks":"To the editor: the parameter-count inconsistency in §3.5.2 versus Table 6 is serious; if the authors cannot supply the actual model definition or code, the efficiency claims should not be trusted. The paper is also below the comparison standard of the journal as it stands. I am recommending major revision because the issues are potentially fixable by correcting the architecture description, adding baselines and ablations, and removing or substantiating the 'state-of-the-art' claim. I would want to see the corrected manuscript before an acceptance decision."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nBottom line: this paper has a real but very incremental idea, and a load-bearing inconsistency between the architecture it describes and the parameter count it advertises. The stress-test note is right. If the final model actually uses 1x1 convs in place of the Dense layers, the MNIST model comes to about 20,622 parameters, not 14,862. The reported 14,862 matches either keeping the original Dense(10) layers or converting to a 5x5 conv (the standard dense-to-conv reshape from Long et al.). The paper says kernel size (1,1) and cites Long et al., so the text is internally inconsistent. This must be fixed before any size claims are credible.\n\nWhat the paper does well: it is a straightforward, honest write-up of a training recipe. The dual-input idea—one submodel on original, one on augmented data, then concatenating—is a reasonable ensembling trick, and combining it with progressive unfreezing is not something I have seen exactly in this form. The author explicitly says results are not compared to other models, which is honest. The latency and throughput measurements are a useful extra, and the model size is genuinely small.\n\nThe soft spots are the usual ones but they matter. No code, no data pipeline, no ablations, no repeated-seed variance. The causal claim that the dual-input design and progressive unfreezing drive the accuracy is untested. \"State-of-the-art\" in the abstract is unsupported; on MNIST, 99% with ~15k parameters is good but not new territory. The CIFAR-10 result (65%) shows limited scalability, which the author acknowledges. The final model's input wiring is ambiguous, and the 1x1 vs 5x5 confusion is central. The reference list is a bit of a grab bag, but nothing that changes the science.\n\nWho is this for? Someone building tiny CNNs for MNIST-like tasks might find the recipe worth trying. It is not a methodological breakthrough. With code, a corrected architecture description, and a couple of baselines (e.g., a single-branch CNN with the same param count and augmentation), it could become a citable empirical note. Without those, it is an interesting but unverified write-up.\n\nMy recommendation: I would send it to review, mainly to force the author to release code and resolve the parameter discrepancy. If the arithmetic is explained (likely a typo of 1x1 instead of 5x5), the paper still needs baselines and ablations for the core claim. As is, I would not accept it.","headline":"A small-CNN recipe worth a look, but the advertised 14,862-parameter count contradicts the architecture as described, and the 'state-of-the-art' label has no baseline to support it.","tokens_in":12607,"tokens_out":5572,"would_cite":false,"duration_ms":47940,"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 two-branch training recipe produces a 14,862-parameter CNN that reaches 99% accuracy on MNIST.","keywords":["lightweight CNN","dual-input-output model","progressive unfreezing","transfer learning","MNIST","fashion MNIST","CIFAR-10","model compression"],"falsifier":"Train a single-branch CNN with the same 14,862 parameters and the same random rotation, zoom, brightness, and translation augmentation on MNIST and fashion MNIST. If it reaches 99% and 89% test accuracy, then the reported results do not depend on the dual-input-output design or progressive unfreezing, and the causal claim would be refuted.","tokens_in":11396,"feed_emoji":"⚡","tokens_out":4404,"duration_ms":34337,"temperature":0.7,"pith_summary":"This paper proposes a two-stage recipe for building very small convolutional networks from the ground up, rather than compressing a large pretrained model. First, two identical submodels are trained side by side, one on the original images and one on augmented copies, and their outputs are concatenated. Second, the trained weights are fused into a single model and fine-tuned by progressively unfreezing layers. The author reports that the resulting model contains only 14,862 parameters (0.17 MB) and reaches 99% accuracy on handwritten digit MNIST and 89% on fashion MNIST, suggesting that task-sized networks can match much larger models.","feed_headline":"Tiny CNN with 14,862 parameters scores 99% on MNIST","feed_subtitle":"Two-branch training plus progressive unfreezing yields a 0.17 MB model with high accuracy.","key_machinery":"The load-bearing mechanism is the dual-input-output construction: two identical feature extractors, one fed the original image and one fed a dynamically augmented version, are trained simultaneously and their outputs are concatenated. The dense layers are then converted to 1x1 convolutions so that the pre-trained weights survive the transition to a single-input model, and progressive unfreezing starting from the last layer fine-tunes the fused model. This combination is what the paper credits for the model's small size, robustness, and accuracy.","core_discovery":"The central claim is that dual-input-output training followed by transfer learning with progressive unfreezing produces a unified single-input model that is both extremely small and accurate. The architecture uses two small convolutional submodels with 10 and 20 filters, concatenates their outputs, converts the dense layers into 1x1 convolutions to preserve learned weights, and adds a small fusion head before fine-tuning. The author reports test accuracies of 99% on handwritten digit MNIST, 89% on fashion MNIST, and 65% on CIFAR-10, with 14,862 and 19,622 trainable parameters respectively, and a model size of 0.17 MB (0.21 MB for CIFAR-10).","pith_inferences":["If the causal claim holds, the paper implies a general recipe: instead of training one model on augmented data, train two branches on complementary views and then fuse them, which may be applicable to other domains such as audio or tabular data.","The 1x1 convolution conversion of dense layers suggests the method extends naturally to fully convolutional architectures and could be combined with semantic segmentation heads.","A direct head-to-head test against LeNet-5 with matched parameter counts would sharpen the 'state-of-the-art' claim, which the paper itself does not provide.","The CIFAR-10 result hints at a complexity ceiling: scaling to harder datasets would likely require more filters or depth, and a rule for that scaling is left open as future work."],"forward_implications":["A model small enough to fit in 0.17 MB can reach 99% accuracy on MNIST, which is relevant for on-device and embedded inference.","The same pipeline can be applied to other small-image classification tasks by only changing the input shape, without structural modifications.","The method works without post-training quantization or pruning, so the reported size is the architecture's natural footprint.","On CIFAR-10 the accuracy drops to 65%, which indicates the method's limits on more complex, color-rich datasets.","Progressive unfreezing with a final SGD fine-tune is presented as the step that stabilizes the fused model and prevents overfitting."],"supporting_citations":[{"why":"Supplies the progressive unfreezing technique used in the second training stage.","marker":"Howard and Ruder (2018)"},{"why":"Defines transfer learning, which the fine-tuning stage relies on.","marker":"Pan and Yang (2010)"},{"why":"Motivates the augmentation techniques applied to the second submodel.","marker":"Shorten and Khoshgoftaar (2019)"},{"why":"Provides the dense-to-convolutional conversion used to preserve weights when building the final model.","marker":"Long et al. (2015)"},{"why":"Serves as the baseline comparison for parameter count and accuracy on MNIST.","marker":"LeCun et al. (1998)"},{"why":"Supplies the dropout used in the fusion head.","marker":"Srivastava et al. (2014)"},{"why":"Justifies the k-fold cross-validation in the final training stage.","marker":"Kohavi (1995)"}],"fun_headline_variants":["99% MNIST with a 14,862-parameter CNN","Pocket CNN: 0.17 MB, 99% MNIST","Lightweight CNN hits 99% on MNIST","14K params, 99% MNIST CNN","Tiny CNN: 99% MNIST at 0.17 MB"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The results are attributed to the specific dual-input plus progressive-unfreezing pipeline, but the paper does not compare against a plain small CNN trained with the same augmentation, so the method's unique contribution is not isolated.","fun_headline_variants_meta":{"raw":{"variants":["99% MNIST with a 14,862-parameter CNN","Pocket CNN: 0.17 MB, 99% MNIST","Lightweight CNN hits 99% on MNIST","14K params, 99% MNIST CNN","Tiny CNN: 99% MNIST at 0.17 MB"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000742,"raw_usage":{"total_tokens":3311,"prompt_tokens":947,"completion_tokens":2364,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":563,"completion_tokens_details":{"reasoning_tokens":2275}},"tokens_in":563,"tokens_out":2364,"duration_ms":15839,"temperature":1.0,"reasoning_tokens":2275,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T14:11:29.862855+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Train a single-branch CNN with the same 14,862 parameters and the same random rotation, zoom, brightness, and translation augmentation on MNIST and fashion MNIST. If it reaches 99% and 89% test accuracy, then the reported results do not depend on the dual-input-output design or progressive unfreezing, and the causal claim would be refuted.","supporting_citations":[{"cited_title":"A study of cross-validation and bootstrap for accuracy estimation and model selection","cited_arxiv_id":null,"evidence_quote":"Justifies the k-fold cross-validation in the final training stage."}],"review_version":1}