{"id":"4a1ca065-62f5-4fa1-ab7c-044259dc25db","arxiv_id":"2507.19321","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"SIDE produces sparse prototype-based explanations for image classifiers, cutting explanation size by over 90 percent while keeping accuracy comparable to InfoDisent.","lead":"SIDE is a new method for prototype-based explainable AI that makes explanations much smaller by pruning away most prototype-class connections. It matches the accuracy of prior methods while using far fewer prototypes per prediction, and it works on large-scale datasets like ImageNet.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The reported >90% explanation-size reduction is an artifact of comparing SIDE after explicit pruning (about A prototypes/class) with InfoDisent counted as all 768 dense prototypes; a matched extraction rule could shrink InfoDisent's local size and collapse the headline number.","rationale":"The strongest evidence for the paper is the accuracy retention: SIDE's top-1 numbers are close to or above InfoDisent across Tables 1 and 2, and Figure 8 shows genuinely concentrated weight distributions after ASL training. I am not contesting that the pruned sigmoid head can retain accuracy. But the central claim has two halves: compactness and understandability. The compactness half is at risk because Local Size is defined by the pruning/OCLA mechanism rather than measured under a comparison that holds the explanation-extraction rule fixed. The paper itself shows InfoDisent has a long tail of small activations, so counting every prototype as 'used' inflates the baseline. This concern is more fundamental than the reader's 'compactness is not interpretability' point: until the size reduction is established under a fair protocol, there is no compactness advantage to serve as a proxy for interpretability. I therefore keep the CONDITIONAL verdict, with the additional explicit condition that the authors report a matched-protocol comparison and, ideally, a human evaluation. Note also the internal inconsistency in Section 5.1: 'over 10x Global Size' vs Table 1 ratios of about 6.6x and 3.9x, which supports the need for corrected reporting.","tokens_in":12993,"tokens_out":10624,"duration_ms":111064,"concrete_test":"Recompute Table 2 under a symmetric explanation-extraction rule. For each test image and predicted class, define the explanation as the smallest prototype subset whose contribution to the class logit reaches 85% (and separately a fixed top-10 subset), for both SIDE and InfoDisent; report Local Size and subset fidelity (how much of the logit/score is retained). If InfoDisent's matched Local Size drops to the same order as SIDE's, the >90% reduction is an artifact of counting all 768 prototypes as the explanation; if it remains an order of magnitude larger, the compactness claim is confirmed.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"SIDE's headline Local Size is not an independent measurement; it is the output of the paper's own explanation-selection protocol. Section 4.3 hard-prunes the Scores Sheet to about A prototypes per class, and Section 4.2's OCLA loss (Eq. 4) is explicitly crafted so that exactly one class exceeds threshold t, so the displayed explanation is confined to that class's surviving prototypes. InfoDisent, by contrast, is reported at Local Size equal to the full prototype count (768 on ImageNet, 2048 on fine-grained sets), even though Figure 1 shows that only 21 of its 768 prototypes activate above 0.2 and its top 4 account for just 30% of the class score. The central quantitative claim thus compares a pruned-by-construction SIDE with an unpruned, unthresholded InfoDisent. If the same extraction rule (threshold at t, or top-k contributions) is applied to both, InfoDisent's effective Local Size could be far below 768, and the '>90% reduction' becomes an accounting artifact rather than a property of the method. The paper's FunnyBirds evaluation (Section 5.3) measures alignment of explanations with model behavior, not human understanding, so the qualitative claim 'substantially enhancing understandability' is not independently supported either.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper introduces Sparse Information Disentanglement for Explainability (SIDE), an extension of InfoDisent for prototypical-part explanations. SIDE decouples the prototype count from the backbone feature dimension via a 1x1 convolution, replaces softmax with sigmoid activations and cross-entropy with asymmetric loss, hard-prunes the Scores Sheet to about A prototypes per class, and adds a calibration stage with a One Correct Label Activation (OCLA) regularizer. Experiments on CUB, Stanford Cars, Stanford Dogs, and ImageNet report top-1 accuracy close to InfoDisent while using far fewer prototypes per prediction, and FunnyBirds results are reported for explanation correctness, completeness, and contrastivity. The central claim is that SIDE matches InfoDisent's accuracy while reducing explanation size by over 90%, thereby substantially enhancing the understandability of prototype-based explanations.","tokens_in":13348,"tokens_out":6855,"duration_ms":61486,"significance":"If the size reduction is measured against a matched baseline, the paper would make a useful contribution: it shows that a sigmoid-based, pruned prototype head can maintain accuracy on ImageNet with SwinV2-S (82.6% vs. 81.4% for InfoDisent) while activating only about 8.6 prototypes per image. The paper is strengthened by extensive ablations, experiments across CNN and ViT backbones, and a FunnyBirds evaluation. The main weaknesses are the asymmetric size comparison, the internal inconsistency of the reported size-reduction factors, and the use of a self-defined and directly optimized metric as evidence for interpretability; these need to be fixed before the headline claims can be accepted.","major_comments":[{"comment":"The claim of \"reductions of over 10x in Global Size\" is not supported by the tables. On CUB with ResNet-50, Global Size drops from 2048 to 310 (6.6x); on DenseNet-121 it drops from 1024 to 266 (3.9x); on ImageNet, SIDE's Global Size (1340 for A=10, 947 for A=7) is larger than InfoDisent's 768. The claim in §5.2 that SIDE explanations are \"over 100x smaller\" than InfoDisent's is also not supported by Table 2, where 768 divided by 8.6 is approximately 89x. Please re-state the quantitative claims so that they match the reported numbers.","section":"§5.1, Tables 1–2"},{"comment":"The Local Size comparison is asymmetric. InfoDisent is assigned a Local Size equal to its total prototype count (2048, 1024, or 768), even though Figure 1 states that for an example ImageNet prediction only 21 of 768 prototypes are active above 0.2. SIDE's Local Size is the output of its own hard-pruning (§4.3) and OCLA thresholding (§4.2) protocol, which by construction keeps about A prototypes per class and forces exactly one class above threshold t. The reported \">300x\" Local Size reduction therefore conflates the method's effect with the extraction rule. The authors should apply the same thresholding or top-k extraction to InfoDisent and report matched Local Size values; until then, the headline reduction is not an independent property of SIDE.","section":"§5.1–§5.2, Tables 1–2, Figure 1"},{"comment":"The OCLA metric is defined in Eq. (3), and then the OCLA loss (Eq. 4) is added in the calibration stage to optimize exactly that metric; reporting improved OCLA and smaller Local Size as evidence of interpretability is therefore circular with respect to the design objective. In addition, the FunnyBirds evaluation in §5.3 measures alignment of explanations with model behavior (correctness, completeness, contrastivity), not human understanding, so the abstract's claim that SIDE \"substantially enhances the understandability\" of prototype-based explanations is not supported by the presented evidence. Please either soften the claim or add a human-subject study that tests whether the compact explanations are actually easier for users to understand.","section":"§4.2–§4.3, §5.3"}],"minor_comments":[{"comment":"The text reports 82.8% top-1 accuracy for SIDE (A=10) and a 0.6 pp gap to the 83.4% baseline, but Table 2 shows 82.6% accuracy, which is an 0.8 pp gap. Please reconcile the numbers.","section":"Abstract, §5.2, Table 2"},{"comment":"The typeset formula for OCLA does not match the prose definition (\"exactly one class is activated\"). As printed, the expression appears to check only that the ground-truth class is activated. Please rewrite the formula with an explicit logical conjunction, for example \\mathbb{1}\\left[\\left(\\sum_j \\mathbb{1}(P_{ij} > t)\\right) = 1 \\wedge P_{iY_i} > t\\right].","section":"§4.2, Eq. (3)"},{"comment":"The word \"completeness\" is misspelled as \"completness\" twice in §5.3, and \"Densnet-121\" appears in the experimental setup text instead of \"DenseNet-121.\"","section":"§5.3, experimental setup"},{"comment":"The tables report single-run accuracies with no standard deviations or number of seeds. Given that several comparisons are within a few tenths of a percentage point (e.g., 82.0 vs. 82.6 on CUB DenseNet-121), please report variability across multiple runs.","section":"Tables 1 and 2"},{"comment":"The paper states \"we make code available\" but does not provide a URL. Please include a repository link.","section":"Code availability"},{"comment":"The limitations section mentions the multi-stage training procedure; it would also be useful to note the sensitivity of the headline compactness result to the choice of A, t, and the OCLA coefficient λ, since these parameters directly determine the reported Local Size and Global Size.","section":"Limitations"}],"recommendation":"major_revision","confidential_remarks":"The paper is within scope and the core idea is plausible. My main concern is that the headline size-reduction claim rests on comparing SIDE's pruned, thresholded output with InfoDisent's full unpruned prototype set; the authors already show in Figure 1 that InfoDisent's effective active set is much smaller than 768. This is fixable with a matched-protocol experiment, which is why I recommend major revision rather than rejection. I would also ask the editor to ensure the abstract's \"substantially enhancing understandability\" wording is either supported by a human study or tempered, since the current evidence is about model-behavior alignment rather than human comprehension."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"First, what you should know: SIDE is a real engineering contribution to prototype-based XAI, but its headline claim of reducing explanation size by over 90% is not as clean as it looks. The comparison counts SIDE after explicit hard pruning (about A prototypes per class) against InfoDisent counted as all 768 or 2048 prototypes. The stress-test note is right: if you applied a similar thresholded or top-k extraction rule to InfoDisent, its effective local size would drop substantially, and the gap would narrow a lot.\n\nWhat's genuinely new: swapping softmax for sigmoid activations with Asymmetric Loss, decoupling prototype count from backbone channels via expansion, and the four-stage training recipe (pretrain, hard prune, fine-tune, calibrate) is a plausible and effective way to get sparse prototype explanations. The accuracy numbers are close to InfoDisent across CUB, Cars, Dogs, and ImageNet, and that part is credible. The ablations on pruning, fine-tuning, and calibration are useful and show the method is not just a black-box trick. Those components are worth building on.\n\nSoft spots, in proportion. First, the local-size comparison is apples-to-oranges: SIDE's explanation is defined by its own threshold and pruning, while InfoDisent's is the full dense weight set. Second, the numbers are sloppy: the text says \"over 10x\" global size reduction and 82.8% ImageNet accuracy, but Table 1 shows 6.6x at best and Table 2 shows 82.6%. No error bars anywhere. Third, the OCLA metric is directly optimized by the OCLA loss and then reported as evidence of interpretability, which is circular. And FunnyBirds measures alignment with model behavior, not human understanding, so \"substantially enhances understandability\" is not actually supported. Fourth, the paper claims code and supplementary details are available, but none are provided in this version.\n\nOverall, this is a serious and competent paper. The core idea is not a huge conceptual leap, but the training recipe works and the fine-grained and ImageNet results are useful for the XAI community. A careful reviewer should engage with it, but the authors need to fix the evaluation protocol, report variance, and either provide code or temper the interpretability claim.\n\nRecommend: send it to peer review, with expectations of major revision.","headline":"A solid engineering improvement to prototype-based XAI, but the headline explanation-size reduction is partly an artifact of comparing pruned SIDE to unpruned InfoDisent; deserves review with major revision.","tokens_in":13849,"tokens_out":3087,"would_cite":false,"duration_ms":30387,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"SIDE shows that prototype-based explanations of image classifiers can be made sparse — about nine prototypes per ImageNet prediction instead of 768 — without sacrificing accuracy.","keywords":["prototypical parts","sparse explanations","interpretable machine learning","multi-label classification","asymmetric loss","hard pruning","fine-grained classification","ImageNet-scale interpretability"],"falsifier":"A user study in which participants are asked to predict or verify a model's decision from SIDE's 8.6-prototype explanations and from InfoDisent's full-prototype explanations; if comprehension is not better for the compact explanations, the central interpretability claim fails.","tokens_in":12819,"feed_emoji":"🧩","tokens_out":6575,"duration_ms":60346,"temperature":0.7,"pith_summary":"The paper introduces SIDE, a method for making prototype-based explanations of deep image classifiers drastically more compact. On ImageNet with a SwinV2-S backbone, SIDE reaches 82.6% top-1 accuracy while using on average 8.6 prototypes per prediction, compared with InfoDisent's 81.4% accuracy and 768 prototypes. SIDE achieves this by replacing the softmax head with independent sigmoid activations, training with an asymmetric multi-label loss, and then hard-pruning the class-to-prototype weights, followed by fine-tuning and a calibration stage. The paper argues that this sparse, disentangled prototype space is faithful to model behavior, reporting improved or matched scores on the FunnyBirds interpretability benchmark. If correct, the result shows that compact explanations need not come at the cost of accuracy on large-scale tasks.","feed_headline":"SIDE explains ImageNet predictions with 9 prototypes instead of 768","feed_subtitle":"A new pruning scheme matches prototype-model accuracy while shrinking explanation size by over 90 percent.","key_machinery":"The central mechanism is a sparse, non-negative Scores Sheet: a class-to-prototype weight matrix $W$ whose entries are passed through ReLU so that zeroed weights stay zero, combined with independent sigmoid activations per class and an Asymmetric Loss that suppresses gradient signals from easy negatives. Together these produce emergent sparsity during training without an explicit sparsity penalty. Hard pruning then retains only the top $A$ weights per class, and a one-correct-label-activation (OCLA) regularizer pushes the model to activate exactly one class above threshold $t$, yielding compact explanations. A prototype expansion via a $1\\times1$ convolution decouples the number of prototypes from the backbone's native channel count, so the method works on transformers with moderate channel widths.","core_discovery":"SIDE's central discovery is that prototype-based explanations can be made sparse without sacrificing predictive accuracy by treating classification as a multi-label problem rather than a single-label one. The method disentangles backbone feature maps into channels, upsamples the channel dimension with a $1\\times1$ convolution, maps each channel to a prototype activation through the sparse pooling operator, and computes class scores with a non-negative ReLU linear layer followed by sigmoid. Training uses Asymmetric Loss, which naturally drives most prototype-class weights to zero during pretraining; a hard pruning step then keeps only the top $k$ weights per class; fine-tuning recovers accuracy; and a final OCLA calibration stage ensures that usually exactly one class is activated. The outcome is a model whose local explanation for a prediction is a handful of prototypes (8.6 on ImageNet at $A=10$, 6.9 at $A=7$) rather than the full prototype set, while matching or exceeding InfoDisent's accuracy.","pith_inferences":["The paper does not test whether fewer prototypes actually improve human comprehension; a direct behavioral study is the natural next step.","If the sparsity effect comes from the combination of Asymmetric Loss, ReLU activations, and pruning rather than from SIDE-specific wiring, the same recipe should transfer to other prototype-based methods.","Because sigmoid class probabilities are independent, SIDE's explanations could extend naturally to multi-label and open-set recognition, where softmax heads tend to mask ambiguity.","The OCLA metric itself could serve as a simple, threshold-based interpretability score for any prototype-based model, independent of SIDE's training procedure."],"forward_implications":["On ImageNet with a SwinV2-S backbone, SIDE matches InfoDisent's accuracy (82.6% vs. 81.4%) while reducing the average local explanation from 768 prototypes to 8.6, and to 6.9 with stronger pruning.","Across CUB, Stanford Cars, and Stanford Dogs, SIDE keeps accuracy within 0.7 points of InfoDisent while cutting global prototype counts by more than 10x and local counts by more than 300x.","The four-stage training schedule recovers accuracy after pruning, and the OCLA calibration stage restores the model's tendency to activate exactly one class, keeping explanations focused.","SIDE's explanations outperform InfoDisent's on the FunnyBirds correctness and completeness dimensions, indicating that the sparse prototypes align with the model's actual decision evidence."],"supporting_citations":[{"why":"Introduces InfoDisent, the frozen-backbone information-disentanglement method that SIDE extends and the main accuracy baseline.","marker":"[44]"},{"why":"Supplies the Asymmetric Loss that produces emergent sparsity and stable multi-label training in SIDE.","marker":"[4]"},{"why":"Provides the FunnyBirds benchmark used to measure correctness, completeness, and contrastivity of SIDE's explanations.","marker":"[18]"},{"why":"ImageNet is the large-scale benchmark where SIDE must scale and where the 8.6-prototype claim is demonstrated.","marker":"[36]"},{"why":"SwinV2 is the backbone on which SIDE reports its ImageNet accuracy and explanation-size results.","marker":"[27]"},{"why":"ProtoPNet introduced prototypical-parts explanations, the paradigm that SIDE's sparse disentanglement builds on.","marker":"[10]"}],"fun_headline_variants":["SIDE cuts ImageNet explanations to 9 prototypes","Explainable AI: 9 prototypes, not 768","SIDE: 90% fewer prototypes, same accuracy","Sparse prototypes: 9 instead of 768 for ImageNet","From 768 prototypes to 9: SIDE simplifies AI"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The paper's central interpretability claim rests on treating explanation compactness—fewer prototypes per prediction—as a proxy for human understanding; no human study or external benchmark ties the reduced prototype count to actually better comprehension.","fun_headline_variants_meta":{"raw":{"variants":["SIDE cuts ImageNet explanations to 9 prototypes","Explainable AI: 9 prototypes, not 768","SIDE: 90% fewer prototypes, same accuracy","Sparse prototypes: 9 instead of 768 for ImageNet","From 768 prototypes to 9: SIDE simplifies AI"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000599,"raw_usage":{"total_tokens":2784,"prompt_tokens":911,"completion_tokens":1873,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":527,"completion_tokens_details":{"reasoning_tokens":1790}},"tokens_in":527,"tokens_out":1873,"duration_ms":13139,"temperature":1.0,"reasoning_tokens":1790,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T17:54:32.401154+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"A user study in which participants are asked to predict or verify a model's decision from SIDE's 8.6-prototype explanations and from InfoDisent's full-prototype explanations; if comprehension is not better for the compact explanations, the central interpretability claim fails.","supporting_citations":[{"cited_title":"Asymmetric loss for multi-label classification, 2021","cited_arxiv_id":null,"evidence_quote":"Supplies the Asymmetric Loss that produces emergent sparsity and stable multi-label training in SIDE."},{"cited_title":"Fun- nybirds: A synthetic vision dataset for a part-based analysis of explainable ai methods","cited_arxiv_id":null,"evidence_quote":"Provides the FunnyBirds benchmark used to measure correctness, completeness, and contrastivity of SIDE's explanations."},{"cited_title":"Swin transformer v2: Scaling up capacity and resolution, 2022","cited_arxiv_id":null,"evidence_quote":"SwinV2 is the backbone on which SIDE reports its ImageNet accuracy and explanation-size results."},{"cited_title":"This looks like that: Deep learning for interpretable image recognition, 2019","cited_arxiv_id":null,"evidence_quote":"ProtoPNet introduced prototypical-parts explanations, the paradigm that SIDE's sparse disentanglement builds on."}],"review_version":2}