{"id":"55ed953e-367c-40bf-9ee3-55f0fc5d3787","arxiv_id":"2411.18410","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":3.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"Stitching persistence images computed by topological data analysis into CNN inputs improves noisy-MNIST accuracy for small training sets, e.g., from 26% to 69% at 100 clean training images.","lead":"The authors combined MNIST handwritten digit images with topological data analysis features, called persistence images, and fed both into a convolutional neural network. On small training sets with noisy test data, this \"Vector Stitching\" approach scored much higher accuracy than a plain CNN.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The V-S accuracy gain is not isolated from the binarization/denoising step in §2.1; without a binarized-image control, the claim that TDA (persistent homology) causes the improvement is unsupported.","rationale":"The reader's verdict is CONDITIONAL, and my stress-test does not move it. The paper's tables are internally consistent and the repeated 10-iteration protocol is a positive feature, but the central empirical claim requires attributing the V-S improvement to topological information. The most concrete threat is the binarization step in §2.1: persistence images are built from thresholded binary images, so the V-S model receives a substantially denoised representation that the RAW baseline never sees. A binarized-input control is missing, and this is not a matter of outside-consensus disagreement; it is a direct confound that could fully explain the accuracy gap. The reader's weakest_assumption identified multiple possible confounds, including binarization and input tensor size; I agree with that direction but single out binarization as the most likely and most readily testable mechanism. If the binarized control matches V-S, the paper's TDA-specific claim fails; if it does not, the remaining gap still needs an input-size control before persistent homology can be credited. I therefore keep the reader's CONDITIONAL verdict unchanged rather than escalating to REJECT, because the evidence is insufficient but not contradictory, and the missing controls are well-specified and feasible to run.","tokens_in":10091,"tokens_out":4066,"duration_ms":42079,"concrete_test":"Retrain the RAW baseline on binarized versions of the training images produced by the exact threshold from §2.1 (pixel value > 0.4) and evaluate on the same noisy test sets used in Table 1. If RAW + binarization reaches or approaches V-S accuracy (e.g., roughly 69% at 100 clean training images instead of 26%), the reported advantage is attributable to denoising, not to persistent homology. As a secondary follow-up within the same experiment, replace the 34 PI channels in V-S with 34 constant or random channels of the same shape; if that control also matches V-S, then input dimensionality is the confound. The decisive first check is the binarized-input control.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section 2.1 starts the TDA pipeline with Binarization: each pixel is assigned 1 if its value exceeds a fixed threshold (0.4), otherwise 0. Every topological feature in the persistence images is computed from this binary image, so the PI channels delivered to the Vector-Stitching model are, at their root, a thresholded and noise-clipped view of the input. In the clean-training/noisy-test protocol of Table 1, the RAW model sees raw noisy test pixels, whereas V-S receives those same raw pixels plus topological summaries of a binarized image. The reported gap (e.g., 69.3% vs 26.0% at 100 clean training images) could therefore be explained by the thresholding operation alone: binarization removes much of the white noise and many salt-and-pepper outliers, acting as a strong denoising preprocessing step. Section 3.2 states only that the three models have 'as similar a total number of parameters as possible' and does not specify architectures, optimizers, training schedules, or the exact input tensor shape; there is also no control for the much larger input dimensionality of V-S. The central claim that TDA improves neural network performance requires isolating the topological summary from the binarization and from input size. The information-theoretic discussion in §5.1 and §5.3 does not supply this isolation: the conditioning intuition is not a valid guarantee for finite-sample training, and the Data Processing Inequality limitation acknowledged in §5.3 concerns information bounds, not this experimental confound.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes 'Vector Stitching' (V-S), a method that concatenates raw grayscale MNIST images with persistence images computed by a topological data analysis (TDA) pipeline, and compares three CNNs: a raw-image model, a TDA-only model, and the V-S model. Experiments are run on small clean or 90/10 clean/noisy training sets (100 to 1000 images) with noisy test sets, repeated 10 times. The reported results in Tables 1 and 2 show that V-S substantially outperforms the other two models, and the paper interprets this through an information-theoretic discussion in Section 5.","tokens_in":10398,"tokens_out":3767,"duration_ms":38777,"significance":"If the reported empirical advantage is robust, the paper provides a useful demonstration that TDA-derived features can improve CNN classification in low-data, noisy regimes. The authors report means, standard deviations, and confidence intervals over 10 repetitions, and they include an honest limitations subsection. However, the current evidence is a single dataset with one noise setting, the architectures and hyperparameters are not specified, and no control isolates the topological contribution from the binarization/denoising preprocessing. The significance is therefore promising but not yet established at the level claimed in the abstract and introduction.","major_comments":[{"comment":"The comparison does not isolate the topological contribution from the binarization step. The persistence images are computed exclusively from the binarized image B defined by the fixed threshold 0.4, so the V-S model receives the raw noisy image together with summaries of a thresholded/denoised image. A binarized-image-only control, or a raw-plus-binarized control, is necessary; without it, the large V-S advantage (e.g., 0.693 vs 0.260 at 100 clean training images in Table 1) could be caused by threshold denoising rather than by persistent homology. Please add such control models and report their accuracies.","section":"§2.1 and Table 1"},{"comment":"The three models are not specified beyond having 'as similar a total number of parameters as possible.' Without layer counts, kernel sizes, padding, optimizers, learning rates, batch sizes, epochs, and any regularization or augmentation, the reader cannot assess capacity matching or reproduce the experiments. The RAW model's accuracy in Table 1 decreases from 0.260 at 100 images to 0.167 at 1000 images, which is suspicious and suggests training instability or unmatched hyperparameters rather than a matched comparison. Please provide complete model definitions and training details for all three networks.","section":"§3.2"},{"comment":"The experimental evidence is restricted to MNIST under one unspecified noise setting. The paper's general claims that TDA improves neural network performance, and that the information-theoretic interpretation explains this improvement, are not supported beyond this single benchmark. The authors should either substantially restrict the scope of the claims or add at least one additional dataset and a small ablation over noise levels (white-noise variance and salt-and-pepper density are not reported anywhere).","section":"§3.1 and §4"},{"comment":"The information-theoretic discussion is not a valid derivation of the empirical gain. The 'conditioning theorem' is invoked as 'adding information cannot hurt model performance,' but conditioning reduces entropy in expectation and does not guarantee that a computed summary improves finite-sample classifier accuracy. Since Section 5.3 acknowledges the Data Processing Inequality, Section 5 should be presented explicitly as qualitative interpretation rather than as an explanation of the V-S advantage.","section":"§5.1 and §5.3"}],"minor_comments":[{"comment":"In the definition of an abstract simplicial complex, 'all of its subsets belong to V' should read 'belong to F'; the vertices are V, while the subsets of a simplex are faces in the family F.","section":"§1.1"},{"comment":"The text says that the interval [a, a+1] 'corresponds to a 1-dimensional standard simplex'; this is an elementary interval/cube, not a simplex. Please correct the wording.","section":"§1.3"},{"comment":"Item 2 in the caption is called a 'Persistence diagram,' but the pipeline in Section 2.1 produces a persistence image after vectorization; the caption should use the correct term.","section":"Figure 8 caption"},{"comment":"The noise generation is not reproducible: the text does not specify the white-noise distribution/variance or the salt-and-pepper density, nor how the 10 pseudorandom training/test selections were seeded.","section":"§3.1"},{"comment":"The statement that the pipeline generates a persistence image consisting of 34 layers would be easier to verify if the authors stated how the H0 and H1 channels are ordered and combined for the 17 filtrations.","section":"§2.1"}],"recommendation":"major_revision","confidential_remarks":"This reads like a preliminary or workshop-level study: the central idea is clear and the reported effect is large, but the single dataset, missing model details, and absent binarization control make the current submission unsuitable for a fully archival journal. I recommend major revision rather than rejection because the central empirical claim is testable and could be made convincing with additional controls and complete experimental specifications. The information-theoretic framing should be softened or removed from the claims."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"You should know two things about this one. The empirical result is real in the narrow sense: the reported tables are internally consistent, and the Vector-Stitching model does much better than raw and TDA-only baselines on noisy MNIST with small training sets. But the paper's central claim—that the improvement comes from topological information—is not established. The stress-test note is right: the pipeline in Section 2.1 starts with binarization at a fixed threshold, and the persistence images are built entirely from that binarized image. So the V-S model receives raw noisy pixels plus topological summaries of a thresholded, noise-clipped version of the same pixels. Without a control that stitches raw pixels to binarized pixels (or to any other denoised input), you cannot attribute the gain to persistent homology. Binarization alone could plausibly produce most of the effect.\n\nWhat is actually new is modest: concatenating raw images with 34 persistence-image channels is a direct application of known TDA vectorization, in the same line as Hofer et al. (2018) which the paper cites. The novelty is the particular configuration and the clean-training/noisy-test evaluation on MNIST. What the paper does well is describe the pipeline concretely, report means, standard deviations, and confidence intervals across 10 runs, and include an honest limitations section—the Data Processing Inequality discussion and the note that gains may shrink with larger networks show the authors are not overselling. The tables are credible.\n\nThe soft spots are significant, though. Beyond the binarization confound, the three models are only described as having \"as similar a total number of parameters as possible.\" Architecture, optimizer, training schedule, and input tensor shapes are not given. No random or constant control channels are tested. The evidence is one benchmark, MNIST, with test sets of only 100 images. Noise parameters are not specified. Without code or architecture details, the experiment cannot be reproduced, and the specificity of the claim to topology cannot be checked. The information-theoretic discussion in Section 5 is qualitative and does not supply the missing isolation.\n\nWho is this for? Researchers in the TDA-plus-deep-learning subfield who are looking for low-cost input augmentation ideas. They should read it as a suggestive benchmark observation, not a demonstration of mechanism. I would not cite it in the next year, but I would send it to a serious referee. The confounds are fixable—add binarized-image and random-channel controls, report architectures, release code—and the question is worth answering. Right now the paper is a good working paper, not a finished claim.","headline":"A plausible and clearly reported MNIST result, but the TDA-specific claim is confounded by binarization and missing control channels.","tokens_in":10927,"tokens_out":1543,"would_cite":false,"duration_ms":15611,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["68T07","55-08","68P30","55-04"],"pacs":[],"model":"deepseek-v4-flash","headline":"By stitching raw MNIST images with persistence images from topological data analysis, small CNNs reach 70-90% accuracy on noisy test sets where raw-image models stay near 17-26%.","keywords":["Topological Data Analysis","Persistent Homology","Persistence Images","Vector Stitching","Convolutional Neural Networks","MNIST","small-sample image classification","noisy test data"],"falsifier":"Re-run the Vector Stitching experiment with the persistence-image tensor replaced by a random or constant tensor of the same shape, and with a binarized-only raw channel as a second control; if accuracy remains near 80-90% on the 1000-image clean training set, the gain is not specifically topological.","tokens_in":9889,"feed_emoji":"🧩","tokens_out":6576,"duration_ms":54838,"temperature":0.7,"pith_summary":"This paper introduces Vector Stitching, a way to feed convolutional networks both the raw pixels of an image and a topological summary of it. The summary is a persistence image built from persistent homology, and the two inputs are concatenated before the first convolutional layer. On MNIST digits corrupted with white and salt-and-pepper noise, the stitched model reaches roughly 70% accuracy with only 100 clean training images and 82% with 1000, whereas the model trained on raw pixels alone reaches 26% and 17% respectively. The paper interprets this as evidence that topological features carry information ordinary CNNs leave unused, and that preserving it helps most when data are scarce and test inputs are noisy.","feed_headline":"Topology stitching lifts noisy MNIST accuracy from 26% to 70%","feed_subtitle":"Persistence-image layers let small CNNs classify digits from just 100 training samples where raw pixels alone fail.","key_machinery":"The load-bearing object is the persistence image: a finite-dimensional rasterization of a persistence diagram in which each birth-death point is replaced by a Gaussian blob, weighted, and discretized onto a pixel grid. The paper computes these images from cubical complexes over binarized MNIST digits, using 17 filtrations (8 height directions and 9 radial centers), and takes both H0 and H1 features, producing a 34-layer tensor. Vector Stitching concatenates this tensor with the raw image, so the network sees pixel detail and global shape persistence in the same input. The mechanism proposed is that conditioning the network on more of the data's structure cannot reduce the information available for inference; in practice it gives small models a large accuracy boost.","core_discovery":"The central discovery is that adding a topological channel to a CNN's input improves classification far beyond what either channel achieves alone. A CNN shown only persistence images reaches about 30% on the hardest protocol; a CNN shown only raw pixels collapses to between 17% and 26%; the stitched model reaches 70-82% on clean-trained sets and 90% when trained on a 90/10 clean/noisy mixture. The topological models also reach high accuracy within the first epochs, suggesting the extra channel accelerates learning rather than merely enlarging the model.","pith_inferences":["The reported gap may partly come from the binarization step removing salt-and-pepper noise before topology is computed; the paper does not isolate that effect from the topological summary itself.","A direct causal test would replace the persistence-image tensor with a random or constant tensor of identical shape in the stitched model; equal accuracy would indicate the advantage is not specifically topological.","If the information-theoretic reading is right, the method should transfer to other low-signal domains, such as medical imaging or time series, where objects are characterized more by shape and connectivity than by texture.","The 17-filtration, 34-layer choice is heuristic, so a natural extension is to ablate the number and type of filtrations and measure accuracy against persistence-image resolution."],"forward_implications":["If the central claim holds, hybrid TDA-plus-pixels inputs are a practical way to keep small CNNs accurate on noisy images when labeled data are scarce.","Topological features alone are not enough: the TDA-only model saturates well below the stitched model, so the benefit comes from combining topology with raw pixels rather than replacing them.","The early-epoch advantage suggests topological channels act as a strong prior for low-data training, reducing the number of gradient steps needed to reach a given accuracy.","The authors' own limitation note implies the gain may shrink as networks grow or as datasets carry more information; the method's advantage is most clearly demonstrated in the small-data, high-noise regime."],"supporting_citations":[{"why":"supplies the foundational definitions of simplicial complexes and the Vietoris-Rips and Cech constructions used in the TDA pipeline.","marker":"[1]"},{"why":"provides the motivating claim that networks trained on raw data underuse topological properties of images.","marker":"[5]"},{"why":"defines persistence images, the vector representation that the pipeline turns persistence diagrams into.","marker":"[6]"},{"why":"supports the stability of topological vectorizations used to justify persistence-based features.","marker":"[7]"},{"why":"prior demonstration that combining TDA with neural networks improves training on noisy or sparse data.","marker":"[8]"},{"why":"supplies the MNIST benchmark and its standard train/test split used throughout the experiments.","marker":"[19]"},{"why":"provides the software implementation used to compute filtrations and persistence images.","marker":"[21]"},{"why":"defines the height and radial filtration variants used to generate the 17 topological channels.","marker":"[23]"}],"fun_headline_variants":["Adding topology data to CNNs lifts MNIST accuracy from 26% to 70%","Stitching topological features into CNNs boosts accuracy to 70%","Topological data enriches CNN input, raising accuracy to 70%","CNN plus topology achieves 70% where pixels alone hit 26%"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The comparison assumes the three networks are matched closely enough in architecture and training that any accuracy gap comes from the topological information, rather than from the larger stitched input or from the binarization step that filters noise.","fun_headline_variants_meta":{"raw":{"variants":["Adding topology data to CNNs lifts MNIST accuracy from 26% to 70%","Stitching topological features into CNNs boosts accuracy to 70%","Topological data enriches CNN input, raising accuracy to 70%","CNN plus topology achieves 70% where pixels alone hit 26%"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000674,"raw_usage":{"total_tokens":3062,"prompt_tokens":930,"completion_tokens":2132,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":546,"completion_tokens_details":{"reasoning_tokens":2050}},"tokens_in":546,"tokens_out":2132,"duration_ms":15531,"temperature":1.0,"reasoning_tokens":2050,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T11:13:23.104099+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Re-run the Vector Stitching experiment with the persistence-image tensor replaced by a random or constant tensor of the same shape, and with a binarized-only raw channel as a second control; if accuracy remains near 80-90% on the 1000-image clean training set, the gain is not specifically topological.","supporting_citations":[{"cited_title":"Chazal, B","cited_arxiv_id":null,"evidence_quote":"supplies the foundational definitions of simplicial complexes and the Vietoris-Rips and Cech constructions used in the TDA pipeline."},{"cited_title":"McGuire, S","cited_arxiv_id":null,"evidence_quote":"provides the motivating claim that networks trained on raw data underuse topological properties of images."},{"cited_title":"Adams, T","cited_arxiv_id":null,"evidence_quote":"defines persistence images, the vector representation that the pipeline turns persistence diagrams into."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"supplies the MNIST benchmark and its standard train/test split used throughout the experiments."}],"review_version":1}