{"id":"283feb40-7dd2-4b17-9b4a-5f172481a4e4","arxiv_id":"1908.04769","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"Adding an unsupervised Infomax term to a graph neural network gives modest ASD classification gains in some architectures and yields candidate separable brain regions.","lead":"This paper adds a mutual-information term, called Infomax, to a graph neural network that classifies autism spectrum disorder from fMRI brain scans. The addition modestly stabilizes classification accuracy in some architectures and produces visually separable node patterns between ASD and healthy-control children in several brain regions.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Opposite-class negative sampling makes L2 a supervised contrastive loss, not Infomax, so the reported F-score gain and node-separability claim are confounded by label leakage.","rationale":"The most load-bearing concern is an explicit design flaw in the loss function, not an omitted detail. The paper repeatedly calls L2 an \"unsupervised\" Infomax term, but the negative sampling from the opposite class turns it into a supervised contrastive loss that directly encodes label information. This invalidates the central comparison between L and L1 models: any performance gain could stem from the additional label signal rather than from the mutual-information regularization that the paper claims to evaluate. The same flaw introduces circularity into the region-difference detection, since the node embeddings were optimized to separate ASD from HC. The reader's weakest assumption (model selection on test folds) is a legitimate secondary concern, but the label-leakage issue is more severe and is already apparent from the text alone, without needing code or data. The paper cannot support its headline claims as written; a major revision with a correct unsupervised negative-sampling baseline is required. Hence the verdict should move from CONDITIONAL to REJECT.","tokens_in":4993,"tokens_out":6475,"duration_ms":69817,"concrete_test":"Obtain the training code from the authors and verify the negative sampling procedure in the L2 branch. Then rerun the (16,16) architecture with the same hyperparameters but replace opposite-class negatives with DGI-style corrupted graphs (e.g., node-feature shuffling) of the same graph. If the F-score gap between L and L1 disappears or shrinks materially, the claimed Infomax advantage is an artifact of label leakage; if the gap persists with a non-supervised negative distribution, the classification result survives but the current description is still misleading.","verdict_should_be":"REJECT","load_bearing_attack":"Section 2.4 states: \"We randomly sample an instance from the opposite class as the negative sample (X', A').\" Eq. (4) then defines the Infomax term L2 as a discriminator over positive pairs (h_i, s) from the same graph and negative pairs (h'_i, s) from an opposite-class graph. This departs from the Deep Graph Infomax procedure (Ref. [3]) where negatives are generated by a corruption function on the same graph. Because the negative graphs are labeled as the opposite class, L2 is a supervised contrastive term: it teaches the encoder to make node embeddings from the same class agree with the class-specific summary and embeddings from the other class disagree. Consequently, the comparison in Table 1 between L(16,16) and L1(16,16) is not a comparison of \"with vs without Infomax regularization\"; the L model has additional access to class labels through L2. The improvement from 0.57±0.11 to 0.68±0.08 could therefore be due to this extra supervised signal, not to unsupervised MI regularization. The same issue undermines the region-level detection: the node embeddings that produced the Silhouette-based \"separable regions\" were trained with a loss that explicitly pushes ASD and HC node representations apart, making the detection partly circular. This is more fundamental than the hyperparameter-selection ambiguity in Sec. 3.2.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript proposes a graph neural network pipeline for classifying autism spectrum disorder (ASD) versus healthy controls (HC) from task-fMRI parcellated into 148 ROIs, and for detecting regional functional differences between the groups. The encoder is a supervised GraphSAGE network with dense hierarchical pooling; in addition to the classification cross-entropy L1, the model optimizes an Infomax term L2 that is intended to maximize mutual information between node embeddings and a graph summary. Experiments on 75 ASD and 43 HC children report F-scores for L1-only and joint L models under three architectures, and the authors identify 31 ROIs with Silhouette score above 0.1 as showing separable ASD/HC representations, comparing these with GLM z-statistics. The central claims are that adding Infomax improves classification as a regularizer and that the learned node embeddings reveal region-level ASD/HC functional differences.","tokens_in":5246,"tokens_out":4935,"duration_ms":52132,"significance":"If the claims held, the contribution would be a practical demonstration that a local-global mutual-information regularizer improves robustness of GNN-based ASD classification and yields a candidate biomarker-discovery tool from whole-brain fMRI graphs. The pipeline is modular and uses established components (GraphSAGE, differentiable pooling, Deep Graph Infomax), which is a strength for reproducibility of the architecture. The paper is also explicit about its negative-sampling choice, which makes the deviation from DGI auditable. However, the central comparison is currently confounded by label information in the negative sampling and by the absence of statistical testing, so the significance of the reported improvement is not yet established.","major_comments":[{"comment":"The negative graph is sampled from the opposite class: the text states 'We randomly sample an instance from the opposite class as the negative sample (X', A').' With the loss written in Eq. (4), the discriminator is trained to push node embeddings of the opposite class away from the summary of the current class. This makes L2 a supervised contrastive term, not the unsupervised Infomax objective of Deep Graph Infomax (Ref. 3), where negatives are produced by a corruption function on the same graph. Consequently, Table 1's comparison between L and L1 models conflates the effect of the MI regularizer with an additional supervised signal derived from class labels. This is load-bearing for the paper's first central claim.","section":"Section 2.4 and Eq. (4)"},{"comment":"As printed, the negative-sample term is E[1 - log D(h'_i, s)], rather than E[log(1 - D(h'_i, s))]. Maximizing 1 - log D does not correspond to the standard binary cross-entropy or MINE objective and can even encourage D to increase for negative pairs. If this is a typographical error, it must be corrected; as written, Eq. (4) is not a valid MI loss.","section":"Section 2.5, Eq. (4)"},{"comment":"No significance testing is reported, and the claimed benefit of L2 rests on a single architecture: for (8,8), L(8,8) with F-score 0.69±0.05 is not better than L1(8,8) with 0.70±0.06, and for (16), the difference is 0.66 vs 0.63 with overlapping standard deviations. The text also states that 'regularization parameters are adjusted correspondingly to get the best performance' without specifying whether this tuning was done on the held-out folds. If it was, the F-score improvement is not a valid generalization result. A paired test across folds and a nested or pre-registered validation procedure are needed to support the claim that Infomax improves classification.","section":"Section 3.2, Table 1"},{"comment":"The region-level detection is circular and threshold-based. The embeddings used for t-SNE and Silhouette analysis are produced by a model whose objective explicitly separates the two classes at the node level; therefore, observing group separability in these embeddings does not independently confirm functional differences. Additionally, the threshold Silhouette > 0.1 and the visual selection of t-SNE panels are arbitrary, and no multiple-comparison correction is applied over 148 ROIs. The comparison with GLM is qualitative. The region-level claim therefore needs either a held-out embedding protocol or an independent statistical test.","section":"Section 3.2, region-level analysis"}],"minor_comments":[{"comment":"The graph augmentation is described only as 'described in our previous work'; the authors should state the augmentation factor and explain how augmented graphs were split (by subject or by graph) to avoid data leakage between training and test folds.","section":"Section 3.1"},{"comment":"The t-SNE panels in Fig. 2(a) do not include axis labels or scales, and because t-SNE distances are not quantitatively meaningful, the visual impression alone is not sufficient evidence of separability; please clarify how the four displayed regions were selected from the 148 ROIs.","section":"Figure 2"},{"comment":"The notation V = (v_1, ..., v_N)^T in R^{N x D} conflates the vertex set with the node attribute matrix; using X for the attribute matrix would make the encoder input in Eq. (1) consistent with the text.","section":"Section 2.1"}],"recommendation":"major_revision","confidential_remarks":"The main risk is that the reported Infomax benefit is confounded by supervised negative sampling. The fix is within scope: rerun with corruption-based negatives from the same graph, or reframe the contribution as a supervised contrastive regularizer. I would not reject outright because the architecture and experimental setup are clear and the authors disclose the negative-sampling choice."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick read: this is an early preprint that combines GraphSAGE, DHP pooling, and Deep Graph Infomax for ASD versus HC classification on task fMRI, then looks at which brain regions have separable node embeddings. The writing is clean, and the authors are honest that the benefit shows up only for one architecture (the (8,8) case shows no gain). What's new is the specific application and the 31-region map, which is a reasonable exploratory output.\n\nThe soft spots are real. The biggest one is in Sec 2.4: negative samples for the discriminator are explicitly drawn from the opposite class. That makes L2 a supervised contrastive term, not the unsupervised Infomax objective from DGI. So Table 1's comparison of L(16,16) versus L1(16,16) compares a model with extra label access to one without it; the 0.57-to-0.68 F-score gain cannot be attributed to MI regularization. The same issue undermines the region-level separability: you trained node embeddings to push ASD and HC apart, so finding separable regions is partly circular. The authors never flag this.\n\nTwo smaller but related issues: there are no significance tests on the F-score differences, the Silhouette threshold of 0.1 is arbitrary, and the description of model selection ('adjusted correspondingly to get the best performance') leaves open the possibility that test folds guided hyperparameter choice. Eq. (4) also appears to have a typo in the Infomax term. No code or data are provided, so reproducibility is limited.\n\nIf the authors resampled negatives from the same graph using a corruption function, the regularization story could be tested cleanly. As is, the central claim about Infomax is not supported. The paper is still useful as a demonstration of a GNN pipeline for fMRI and as a cautionary example of how negative sampling choices can silently change an unsupervised objective into a supervised one. It deserves a serious referee because the problem is important and the flaw is subtle enough that experts should weigh in, but I'd expect major revision or rejection on the current evidence.","headline":"A clearly written fMRI-GNN paper whose 'Infomax' regularizer turns out to be a supervised contrastive loss because negatives come from the opposite class, so the main classification gain and region-separability result are confounded.","tokens_in":5822,"tokens_out":2445,"would_cite":false,"duration_ms":25586,"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":"Adding an unsupervised Infomax loss to a graph neural network improves ASD classification from fMRI and reveals separable brain-region representations between autism and healthy controls.","keywords":["ASD classification","fMRI","graph neural network","Infomax","mutual information","graph embedding","brain functional differences","biomarker detection"],"falsifier":"Re-run the (16,16) comparison with layer width and regularization chosen on a validation split (or pre-specified), then test on held-out subjects; if the classification-only model reaches the same F-score as the Infomax model, the claimed regularizing benefit collapses. Also, on an independent ASD dataset, check whether the same 31 regions remain separable with Silhouette score above 0.1.","tokens_in":4742,"feed_emoji":"🧠","tokens_out":5837,"duration_ms":55901,"temperature":0.7,"pith_summary":"The paper sets out to show that adding an unsupervised mutual-information term, Infomax, to a graph neural network improves both tasks it is trained for: classifying autism spectrum disorder from fMRI and exposing functional differences between ASD and healthy control brains. The reported result is that on a two-layer (16,16) GraphSAGE architecture, the Infomax-augmented loss raises the classification F-score from 0.57 to 0.68 while the classification-only model overfits. The same trained embeddings separate the two groups in 31 brain regions, including prefrontal and cingulate cortex, with Silhouette scores above 0.1. A sympathetic reader would take the central claim to be that graph-level mutual information acts as a regularizer on noisy fMRI embeddings and makes group-level regional differences visible without explicit region labels.","feed_headline":"Infomax lifts ASD brain-scan classification F-score to 0.68","feed_subtitle":"A mutual-information term regularizes noisy fMRI embeddings and flags 31 brain regions separating autism from controls.","key_machinery":"The mechanism is a three-part pipeline: a GraphSAGE encoder with mean-pooling propagation maps each brain graph to node embeddings; DHP hierarchical pooling summarizes them into a graph-level readout; and a bilinear discriminator scores pairs of node embeddings and the summary, with a loss that maximizes mutual information between local and global representations. The training objective combines binary cross-entropy classification loss with the Infomax term, so node embeddings must serve both classification and representation quality. The Infomax term encourages nodes within the same graph to share common signals while pushing embeddings of negative graphs apart.","core_discovery":"The central discovery claimed is that a graph embedding trained with an added Infomax loss regularizes the noisy fMRI representation and makes ASD and healthy-control differences visible at the node level. Specifically, the paper reports that the (16,16) GraphSAGE model with classification loss only gives an F-score of 0.57 and overfits, while the same architecture with the combined loss gives 0.68. The paper also finds that 31 brain regions, including prefrontal and cingulate cortex, have node embeddings with Silhouette score above 0.1, indicating linear separability of the two groups. These separable regions are presented as functional differences between ASD and HC that align with prior literature and could serve as new biomarkers.","pith_inferences":["A testable extension is that the benefit of Infomax should grow with encoder capacity and data noise; comparing the (8,8) and (16,16) results already hints at this, and larger, noisier fMRI cohorts would put it to the test.","The same pipeline could be applied to other neurodevelopmental or psychiatric conditions, with negative samples drawn from the opposite diagnostic group as the paper does.","The biomarker-detection claim would be stronger if the 31 separable regions were compared against multiple established fMRI contrast maps rather than a single GLM baseline."],"forward_implications":["Adding the Infomax loss to a graph neural network can act as a regularizer, preventing overfitting when the encoder becomes deeper or wider.","The learned node embeddings can be used to locate brain regions with functional differences between ASD and healthy controls without requiring region-level labels.","The detected regions, such as prefrontal and cingulate cortex, align with existing ASD biomarker literature and may suggest new biomarkers when applied to other datasets.","Training a graph embedding with a combined classification and mutual-information loss is a viable strategy for other noisy, high-dimensional graph classification problems."],"supporting_citations":[{"why":"prior graph neural network approach to task-fMRI ASD biomarkers that this work extends with Infomax","marker":"[1]"},{"why":"provides the mutual information neural estimation method enabling MI computation in high dimensions","marker":"[2]"},{"why":"introduces Deep Graph Infomax, the local-global MI objective adapted as the L2 loss","marker":"[3]"},{"why":"supplies the GraphSAGE mean-pooling convolution used as the encoder","marker":"[4]"},{"why":"supplies the differentiable hierarchical pooling that produces the graph-level readout","marker":"[5]"},{"why":"reports neural signatures of autism used as prior literature for validating detected regions","marker":"[7]"},{"why":"defines the preprocessing pipeline and reports biological-motion brain responses in autism used as biomarker context","marker":"[8]"},{"why":"provides the cortical parcellation atlas that defines the 148 brain-region nodes","marker":"[9]"}],"fun_headline_variants":["Infomax boosts ASD brain scan classification to 0.68 F-score","Infomax reveals 31 brain regions separating ASD from controls","Mutual information loss sharpens ASD fMRI graph embeddings","Graph neural net with Infomax flags autism brain differences","Infomax regularization improves ASD classification and region detection"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The central comparison assumes that the choice of layer width and regularization strength, described only as 'adjusted correspondingly to get the best performance,' was made without using the held-out test folds; if test performance guided those choices, the Infomax improvement is not a valid generalization result.","fun_headline_variants_meta":{"raw":{"variants":["Infomax boosts ASD brain scan classification to 0.68 F-score","Infomax reveals 31 brain regions separating ASD from controls","Mutual information loss sharpens ASD fMRI graph embeddings","Graph neural net with Infomax flags autism brain differences","Infomax regularization improves ASD classification and region detection"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000637,"raw_usage":{"total_tokens":2947,"prompt_tokens":966,"completion_tokens":1981,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":582,"completion_tokens_details":{"reasoning_tokens":1897}},"tokens_in":582,"tokens_out":1981,"duration_ms":14188,"temperature":1.0,"reasoning_tokens":1897,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T14:17:43.459121+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Re-run the (16,16) comparison with layer width and regularization chosen on a validation split (or pre-specified), then test on held-out subjects; if the classification-only model reaches the same F-score as the Infomax model, the claimed regularizing benefit collapses. Also, on an independent ASD dataset, check whether the same 31 regions remain separable with Silhouette score above 0.1.","supporting_citations":[{"cited_title":"Graph neural network for interpreting task-fmri biomarkers,","cited_arxiv_id":null,"evidence_quote":"prior graph neural network approach to task-fMRI ASD biomarkers that this work extends with Infomax"},{"cited_title":"Mine: mutual information neural estimation,","cited_arxiv_id":null,"evidence_quote":"provides the mutual information neural estimation method enabling MI computation in high dimensions"},{"cited_title":"Deep graph infomax,","cited_arxiv_id":null,"evidence_quote":"introduces Deep Graph Infomax, the local-global MI objective adapted as the L2 loss"},{"cited_title":"Inductive representation learning on large graphs,","cited_arxiv_id":null,"evidence_quote":"supplies the GraphSAGE mean-pooling convolution used as the encoder"},{"cited_title":"Hierarchical graph representation learning with differentiable pooling,","cited_arxiv_id":null,"evidence_quote":"supplies the differentiable hierarchical pooling that produces the graph-level readout"},{"cited_title":"Neural signatures of autism,","cited_arxiv_id":null,"evidence_quote":"reports neural signatures of autism used as prior literature for validating detected regions"},{"cited_title":"Brain responses to biological motion predict treatment outcome in young children with autism,","cited_arxiv_id":null,"evidence_quote":"defines the preprocessing pipeline and reports biological-motion brain responses in autism used as biomarker context"},{"cited_title":"Automatic parcellation of human cortical gyri and sulci using standard anatomical nomenclature,","cited_arxiv_id":null,"evidence_quote":"provides the cortical parcellation atlas that defines the 148 brain-region nodes"}],"review_version":1}