{"id":"3c41f128-4009-471d-bab6-7d2533bae6f5","arxiv_id":"2607.19315","paper_version":2,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"The effective rank of a deep feature map's channel covariance is a label-free per-image richness score that correlates with human complexity and improves data selection for super-resolution and OCR.","lead":"This paper uses the effective rank of a deep feature map's channel covariance as a per-image score of visual richness, computed in one forward pass through a frozen pretrained encoder. It shows the score tracks human complexity judgments (r=0.72) and can guide data pruning for super-resolution and OCR, but not for classification, segmentation, or denoising.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Layer-averaging of ERank is load-bearing and possibly post hoc: corruption probe (Table 1) shows sign flips at deepest layers; no sensitivity analysis over layers/encoders, so the r=0.72 and data-selection gains may not be robust.","rationale":"The reader's weakest assumption precisely identifies the sensitivity of ERank to encoder and layer choice. My reading of the manuscript confirms that the layer-averaging step is load-bearing: the corruption probe (Table 1) shows non-uniform and even opposite behavior across depths, and the text explicitly uses this to justify averaging, which conflicts with the claim that layers were fixed a priori. The paper acknowledges the limitation but provides no sensitivity analysis. The central claim that ERank is a general richness measure with r=0.72 human correlation is therefore not yet secure. The data-selection results also contain an overclaim: the introduction states ERank 'beats random removal across pruning budgets,' but Table 3 shows CLIP ERank underperforms random at the 30% OCR pretraining budget (0.224 vs 0.285), and ResNet ERank underperforms random at multiple budgets. This strengthens the need for caution. However, the paper has real strengths: a clear definition, an appealing interpretation, multiple validation axes, and a coherent explanation of where selection helps. The concerns do not warrant rejection; they warrant a conditional acceptance pending a systematic layer/encoder sensitivity study and correction of the reported inconsistencies. Since the reader already issued a CONDITIONAL verdict, my assessment leaves that verdict unchanged.","tokens_in":10695,"tokens_out":6155,"duration_ms":60490,"concrete_test":"Recompute the IC9600 Pearson correlation and the Table 3 data-selection results for each layer/block independently (ResNet layers 1-4; CLIP blocks 2,5,8,11) and for every non-empty subset of these layers/blocks. If the per-layer r with human complexity varies by more than ±0.15 around 0.72, or if any single layer inverts the ranking (e.g., blur increases ERank at a deepest layer), then the averaged score is not a stable richness measure and the headline result is conditional on a post hoc averaging choice. Also verify whether Table 1's layer labels match the stated fixed layer set.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim — that ERank is a stable, per-sample richness measure computable from a frozen encoder — depends on the specific choice of layers and their averaging. The paper fixes ResNet layers {1,2,3,4} and CLIP blocks {2,5,8,11} (Sec. 4.1) and averages the per-layer ERank scores. Table 1, however, shows that the deepest layers behave oppositely to the shallow ones under blur: ResNet layer 4 has rank-shift +166 while layer 1 has -498; CLIP block 11 has -368 while block 2 has -490. The paper states that these sign flips 'motivate averaging ERank across depths,' but Sec. 4.1 claims the layers were 'fixed a priori' and 'not tuned.' This is a post hoc justification of a load-bearing design choice. If averaging is justified by the corruption probe, then the probe is being used both to motivate and to validate the measure, a circular step. Without per-layer correlations on IC9600, or per-layer data-selection results, the reported r=0.72 could be an artifact of combining layers that have different (perhaps even negative) relationships with human complexity. The paper's own Limitations section concedes that 'results may differ under a systematic layer-selection study,' yet no such study is provided. Additionally, there is an internal inconsistency: Table 1 lists CLIP blocks 2, 6, 11, while Sec. 4.1 specifies blocks 2, 5, 8, 11; ResNet rows show layers 1, 3, 4, not all 1–4. This undermines confidence in the precision of the experimental report. The absence of sensitivity analysis over encoders, layer subsets, or seeds is the weakest point in an otherwise plausible argument.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes the effective rank (ERank) of the channel covariance of an image's deep feature map as a per-sample, label-free measure of visual richness. ERank is computed in a single forward pass through a frozen pretrained encoder (ResNet-18 or CLIP ViT-B/32), by averaging per-layer ERank values over a fixed set of layers. The authors characterize its mathematical properties, show that it correlates with human complexity annotations on IC9600 (r = 0.72), and evaluate it as a data-selection criterion. They report that removing low-ERank samples improves super-resolution and removing high-ERank samples improves OCR, while it does not help classification, segmentation, or denoising. The paper includes a corruption probe, comparisons to unsupervised activation energy (UAE), and a discussion of limitations.","tokens_in":11146,"tokens_out":3380,"duration_ms":35758,"significance":"If the central claim holds, ERank provides a cheap, label-free, per-sample complexity signal with a clear interpretation and plausible applications in data pruning, curriculum learning, and adaptive computation. The paper has several strengths: the measure is simple and computationally efficient; no parameters are fitted to human labels or task outcomes, so the human correlation is a genuine external validation; the authors compare against random removal and report negative results on three tasks, which is informative; and the corruption probe attempts to characterize failure modes. The r = 0.72 correlation with human complexity on IC9600 is a strong quantitative result. However, the validity of the measure as a stable richness ranking depends on the specific choice of layers and their averaging, and the paper does not provide the sensitivity analysis needed to establish that stability. The data-selection claims are also overstated relative to the reported results. These issues are fixable but require additional experiments and careful rewriting.","major_comments":[{"comment":"The layer-averaging choice is load-bearing and appears post hoc. Section 4.1 states the layers were 'fixed a priori' and 'not tuned,' yet Section 4.3 uses Table 1's sign flips (ResNet layer 4 blur +166 vs layer 1 −498; CLIP block 11 −368 vs block 2 −490) to say that these results 'motivate averaging ERank across depths.' This is a post hoc justification of a design choice that directly affects every reported correlation and selection result. Moreover, Table 1's rows do not match Section 4.1: Table 1 lists CLIP blocks 2, 6, 11 and ResNet layers 1, 3, 4, while Section 4.1 specifies blocks 2, 5, 8, 11 and layers 1–4. The paper's own Limitations section concedes that 'results may differ under a systematic layer-selection study,' but no such study is provided. Without per-layer IC9600 correlations or a robustness analysis over layers, encoders, and seeds, the r = 0.72 and the selection gains","section":"Abstract, §4.4, Table 3"},{"comment":"The claim that 'In both tasks ERank beats random removal across pruning budgets' is false as stated. Table 3, IIIT5K OCR pretraining at the 30% budget: ERank (CLIP) gives 0.224 ± 0.067 vs random 0.285 ± 0.027, i.e., worse than random; ResNet-18 ERank is worse than random at every pretraining budget and at most finetune budgets. The abstract and Section 4.4 need to be corrected to report the specific budgets where ERank helps and where it does not, and to avoid the universal 'beats random' phrasing.","section":"Table 3, Figures 3a/3b"},{"comment":"The super-resolution gains over random are numerically small and their statistical significance is not established. For example, at the 10% pretraining budget, random gives 28.850 ± 0.013 and ERank (CLIP) gives 28.874 ± 0.015; at 50%, 28.688 ± 0.004 vs 28.743 ± 0.013. With only three task-model seeds and no paired significance tests, the claim of consistent improvement 'at every budget' is not supported. The authors should report paired differences, confidence intervals, or additional seeds, and temper the conclusion accordingly.","section":"§4.3, Figure 2"}],"minor_comments":[{"comment":"The layer lists in Table 1 and Section 4.1 are inconsistent. Please unify them and ensure the reported corruption-probe rows correspond to the layers actually used in scoring.","section":"Figure 3"},{"comment":"The caption says curves are distinguished only by color; this is inaccessible for grayscale printing and color-blind readers. Use distinct line styles or markers.","section":"§4.3"},{"comment":"The p-values in Table 2 are all reported as '<0.05' or '0.28/0.27'; for correlations on 2000 samples, exact p-values or confidence intervals would be more informative. Also, the p-value for r=0.72 on IC9600 is not given, only 'p<0.05' in the figure caption.","section":"§3.1"},{"comment":"The paragraph 'Properties of ERank' refers to 'the next section' but Section 3.2 does not exist; the properties are presented in Section 3.1 and Appendices A–B. Please fix the cross-reference.","section":"§4.4"},{"comment":"In the 'Where ERank helps' paragraph, the phrase 'the visually rich images are the hard ones' is slightly confusing; consider clarifying that high ERank correlates with clutter/noise in OCR, which is the reason for removing high-ERank samples.","section":"§4.4"}],"recommendation":"major_revision","confidential_remarks":"The core idea is interesting and the human-correlation result is promising, but the paper currently overstates both the stability of the measure and the data-selection results. The layer-averaging issue is the most important: the paper should either provide a sensitivity analysis (per-layer correlations, alternative layer sets, alternative encoders) or substantially weaken the claim that ERank is a stable per-sample richness measure. The internal inconsistency in Table 1 versus Section 4.1 suggests a need for careful checking of all experimental details. I believe these concerns are addressable within the manuscript's scope, hence major revision rather than rejection."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: the per-sample use of effective rank as a richness score is new and the human correlation is the real result; the data-selection gains are small and the paper oversells one of them. The layer averaging is load-bearing and needs sensitivity analysis before I'd use the score.\n\nThe genuinely new thing here is the within-sample ERank: existing work (Vendi, RankMe) uses ERank on a set of samples, whereas this computes it over channels of a single feature map. That is a reasonable move and the paper gives it a fair amount of validation. The r=0.72 correlation with IC9600 human complexity annotations is substantial, and the fact that ERank tracks sharpness/compressibility but not distortion/colorfulness is the right kind of evidence for a richness measure. I also give them credit for reporting where selection does not help — classification, segmentation, denoising — and for explaining why. That boundary is the most useful part of the paper.\n\nThe soft spots are real but not fatal. The biggest is the layer averaging. Table 1 shows sign flips at deep layers under blur, and the text says this motivates averaging, yet Sec. 4.1 says the layers were fixed a priori and not tuned. You can't have both. Either the averaging was a post hoc choice (then the r=0.72 could be partly an artifact of combining layers with different relationships to complexity), or the layers were chosen ahead of time (then the sign-flip explanation is just a rationalization). The paper needs a per-layer correlation table and a sensitivity analysis over layer subsets and encoders. The stress-test note is right on this. There is also an internal inconsistency: Table 1 lists CLIP blocks 2, 6, 11, while Sec. 4.1 says blocks 2, 5, 8, 11; and ResNet rows show layers 1, 3, 4 not all four. That's the kind of sloppiness that undermines confidence in the numbers.\n\nThe other concrete problem is the abstract's claim that ERank 'beats random across pruning budgets.' Table 3 contradicts it at OCR pretraining 30%, where CLIP ERank (0.224) is below random (0.285), and ResNet-18 is worse in both OCR regimes. The claim should be qualified.\n\nThe math is elementary — bounds, invariance, noise monotonicity — but correctly stated, and the appendix makes the connection to Hill numbers and participation ratio cleanly. No code is provided; that should be a requirement.\n\nBottom line: this is a plausible heuristic for data curation, not a field-changer. With a serious referee asking for robustness analysis and a corrected claim, it could be a solid workshop or good-quality conference paper. Worth a serious referee.","headline":"The per-sample use of effective rank as a richness score is new and the r=0.72 human correlation is the real result; the data-selection gains are modest and the paper oversells one of them, with load-bearing layer averaging that needs a sensitivity analysis.","tokens_in":11636,"tokens_out":2347,"would_cite":false,"duration_ms":25756,"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":"The paper claims that the effective rank of the channel covariance of a single feature map, computed in one forward pass through a frozen pretrained encoder, is a per-image, label-free measure of visual richness—and that pruning training da","keywords":["effective rank","image complexity","visual richness","data selection","channel covariance","label-free","super-resolution","OCR"],"falsifier":"Compute ERank on IC9600 using only a single deepest layer (for example, ResNet-18 layer 4), where the paper's own corruption probe shows sign flips under blur: if the correlation with human complexity drops sharply or the super-resolution/OCR pruning gains vanish, then the reported results are an artifact of the specific layer mixture rather than a stable property of ERank.","tokens_in":10620,"feed_emoji":"🖼️","tokens_out":4245,"duration_ms":44285,"temperature":0.7,"pith_summary":"The paper proposes that ERank—the exponential of the Shannon entropy of the normalized eigenvalues of an image's channel covariance at a chosen layer—counts how many decorrelated channel directions the image activates. It is computed with a single forward pass through a frozen pretrained encoder, costs no labels, and ranks images from plain to visually rich. The authors show this ordering tracks human complexity judgments on IC9600 with r = 0.72 and correlates with codec bitrate, sharpness, and edge density. As a data-selection criterion, removing low-ERank samples improves super-resolution on DIV2K, while removing high-ERank samples improves OCR on IIIT5K, beating random pruning at most budgets; it does not help classification, segmentation, or denoising. The takeaway is a cheap richness signal whose usefulness is bounded to tasks where input richness, rather than class structure, spatial layout, or noise, governs task difficulty.","feed_headline":"Effective rank of deep features scores image richness in one pass","feed_subtitle":"One forward pass, no labels: ERank tracks human complexity ratings and steers data pruning for SR and OCR.","key_machinery":"ERank is the exponential Shannon entropy of the normalized eigenvalue spectrum of the channel covariance of a feature map: with p_i the normalized eigenvalues, ERank = exp(-sum_i p_i log p_i). This single number turns a raw feature map into a per-image complexity score that is invariant to isotropic scaling and channel rotations, continuous and differentiable, and exactly tied to pairwise channel correlation. The implementation averages ERank over four fixed layers of each encoder (ResNet-18 layers 1-4, CLIP ViT-B/32 blocks 2,5,8,11); this layer averaging is load-bearing, because the paper's own corruption probe shows sign flips at the deepest layers under blur.","core_discovery":"The central claim is that image richness can be read off the eigenspectrum of a deep feature map's channel covariance. For an image and a layer, the feature map is flattened to a spatial-by-channel matrix, its centered channel covariance is formed, and the normalized eigenvalues define a distribution; exponentiating the Shannon entropy of that distribution gives ERank. This is a smooth, scale- and rotation-invariant relaxation of matrix rank that equals the number of channels when channels are fully decorrelated and collapses to one when they are perfectly correlated. The paper argues that because trained encoders are biased toward low-rank representations, a broad single-image spectrum is i","pith_inferences":["Because ERank is differentiable, a natural extension the authors note but do not test is to use it as a training objective or regularizer that rewards decorrelated channel spectra, not just as a pruning filter.","The layer-dependent sign flips in the corruption probe suggest per-layer ERank profiles could serve as a corruption or domain-shift diagnostic: a mid-layer drop with a deep-layer rise under blur is a signature pattern worth exploiting.","A testable extension is complexity-routed adaptive computation: use ERank to decide how many network blocks an image needs, since plain images should require fewer layers than visually rich ones.","The paper tests compact task models; whether ERank-based selection gains persist for larger models and larger-scale pretraining remains open and would be the next empirical check."],"forward_implications":["If ERank is a valid richness measure, any image dataset can be ordered plain-to-rich with one forward pass per image, no labels, and no extra training.","Data pruning by ERank improves super-resolution when low-ERank (plain) samples are removed and OCR when high-ERank (cluttered) samples are removed, across pruning ratios up to 50%.","The measure's failure to help classification, segmentation, and denoising delimits where per-sample richness is the right selection signal: task difficulty must be driven by input richness.","ERank's r = 0.72 agreement with human complexity annotations suggests it can serve as a cheap substitute or complement for human complexity labeling."],"fun_headline_variants":["ERank scores image richness from one deep-feature pass","Single forward pass: ERank measures visual complexity without labels","ERank's eigenspectrum ranks image complexity, matches humans","Cheap richness signal: ERank helps pick training data for SR and OCR","One-pass ERank: channel diversity as a richness ruler"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The load-bearing premise is that a fixed pretrained encoder and a fixed set of layers, averaged together, rank image richness consistently enough that the ordering transfers across datasets and tasks; if the layer choice or encoder changes, the human correlation and data-selection gains may not persist.","fun_headline_variants_meta":{"raw":{"variants":["ERank scores image richness from one deep-feature pass","Single forward pass: ERank measures visual complexity without labels","ERank's eigenspectrum ranks image complexity, matches humans","Cheap richness signal: ERank helps pick training data for SR and OCR","One-pass ERank: channel diversity as a richness ruler"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000156,"raw_usage":{"total_tokens":1022,"prompt_tokens":678,"completion_tokens":344,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":422,"completion_tokens_details":{"reasoning_tokens":257}},"tokens_in":422,"tokens_out":344,"duration_ms":3866,"temperature":1.0,"reasoning_tokens":257,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-01T12:45:20.868992+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Compute ERank on IC9600 using only a single deepest layer (for example, ResNet-18 layer 4), where the paper's own corruption probe shows sign flips under blur: if the correlation with human complexity drops sharply or the super-resolution/OCR pruning gains vanish, then the reported results are an artifact of the specific layer mixture rather than a stable property of ERank.","supporting_citations":[],"review_version":1}