{"id":"262283a5-1080-4414-95f4-f4c69046007d","arxiv_id":"2504.20078","paper_version":2,"verdict":"REJECT","confidence":"MODERATE","novelty_score":3.0,"correctness_risk":"high","formal_verification":"none","parameter_count":1,"one_line_summary":"ARSVD compresses pretrained CNNs to 10 to 20 percent fewer parameters while reporting F1 scores 0.1 to 0.5 points higher than the uncompressed baseline.","lead":"The paper proposes compressing each layer of a neural network with a different SVD rank, chosen by how spread out the singular values are, measured with spectral entropy. It reports smaller models and similar or slightly higher accuracy on MNIST, CIFAR-10, and CIFAR-100, but the comparison uses the uncompressed models as the baseline.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Table II's 'SVD' rows match uncompressed model parameter counts, so ARSVD is never compared against actual fixed-rank SVD; the central superiority claim is untested.","rationale":"I agree with the reader's weakest-assumption analysis: the Table II rows labeled 'SVD' exactly match the parameter counts of the uncompressed ResNet18, InceptionV3, and VGG16 models, so the paper's own evidence indicates that no fixed-rank SVD baseline was constructed. This is the single most load-bearing problem because the paper's headline contribution is explicitly comparative: ARSVD is claimed to outperform fixed-rank SVD. Without a fixed-rank arm at a comparable compression level, the reported numbers only compare ARSVD to the full pretrained model, which cannot establish the claimed advantage. The issue is concrete, locatable in Section IV, and directly falsifiable by re-running the experiment with a matched fixed-rank baseline. The undisclosed tau and unspecified reshaping are real but secondary; they affect reproducibility, whereas the missing baseline affects validity of the central comparison. Because the reader's REJECT verdict is based on this same flaw and my analysis does not change that assessment, the verdict remains as originally given.","tokens_in":6887,"tokens_out":3811,"duration_ms":41344,"concrete_test":"Re-run the Section IV experiment with an actual fixed-rank SVD baseline. For each dataset and model, choose a single global rank r so that the total parameter count after fixed-rank truncation matches the ARSVD-reported count (e.g., 9.8M parameters for ResNet18 on CIFAR-10), apply the truncation to the pretrained weights, and evaluate the same F1/accuracy metric without fine-tuning. If fixed-rank SVD at equal parameter budget matches or beats ARSVD, the central comparative claim fails; if ARSVD wins, the claim receives its first direct evidence. As a secondary check, verify Table II by loading the standard pretrained checkpoints and counting parameters: if the 'SVD' rows equal the full pretrained counts, the baseline was uncompressed.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's central claim is that ARSVD 'consistently achieves better compression-performance trade-offs than traditional fixed-rank SVD' (Section I). For that claim to hold, the experiments must include a fixed-rank SVD baseline at a comparable compression level. They do not. In Table II, the rows labeled 'SVD' for ResNet18, InceptionV3, and VGG16 are 11.2, 23.9, and 134.3 million parameters, which are exactly the standard parameter counts of the uncompressed pretrained models. No fixed-rank truncation was performed; 'SVD' functions as an uncompressed baseline. Consequently, the reported deltas (e.g., ResNet18 CIFAR-10 F1 80.1 vs 80.4 with 11.2M vs 9.8M parameters) compare ARSVD to the full model, not to a truncated SVD at an equivalent rank. Even if the reported F1 scores are taken at face value, they cannot support the stated advantage over fixed-rank SVD because the fixed-rank condition was never run. The undisclosed threshold tau and the unspecified reshaping of convolutional kernels are additional reproducibility barriers, but they are secondary: correcting them would not supply the missing comparison.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes Adaptive-Rank Singular Value Decomposition (ARSVD), a compression method that selects a per-layer truncation rank k by requiring the partial spectral entropy H(k) to be at least a fraction tau of the total spectral entropy of the layer's weight matrix. The authors claim that, compared with fixed-rank SVD, ARSVD gives better compression-performance trade-offs on ResNet18, InceptionV3, and VGG16 across MNIST, CIFAR-10, and CIFAR-100. The paper reports parameter counts, F1 scores, and inference-time comparisons in Tables II and III and Figures 1 and 2, and provides pseudocode for the algorithm.","tokens_in":7031,"tokens_out":4820,"duration_ms":45073,"significance":"The idea of using spectral entropy to choose a per-layer rank is a plausible and cheap heuristic that could be useful for quick compression of pretrained CNNs, and the paper correctly identifies that fixed-rank truncation ignores layer-wise sensitivity. If the reported accuracy figures were backed by a real fixed-rank SVD baseline and a disclosed experimental protocol, the method would be a simple practical contribution. As submitted, however, the central comparative claim is not tested: the rows labeled 'SVD' in Table II correspond to the uncompressed model sizes, and no fixed-rank truncation experiment is described. The paper also omits the value of the entropy threshold tau, the reshaping procedure for convolutional kernels, the fine-tuning protocol, and code, so the results are not reproducible in their current form.","major_comments":[{"comment":"The parameter counts in the rows labeled 'SVD' (11.2M for ResNet18, 23.9M for InceptionV3, 134.3M for VGG16) are the standard parameter counts of the uncompressed pretrained models, not the counts of fixed-rank truncated SVD models. No rank k, no truncation rule, and no low-rank factor sizes are reported for the SVD baseline. Therefore the comparison in Tables II-III and Figures 1-2 is between ARSVD and the full models, not between ARSVD and fixed-rank SVD. The central claim in Section I that ARSVD 'consistently achieves better compression-performance trade-offs than traditional fixed-rank SVD' has no supporting evidence as written.","section":"Section IV-A, Table II"},{"comment":"The entropy threshold tau is a free parameter of the method, but its value is never disclosed in Section IV, and no sensitivity analysis is given. Without tau, the reported compressed parameter counts cannot be reproduced, and the reader cannot tell whether the method's behavior is sensitive to this choice. The paper should report the tau values used for each model/dataset, the resulting per-layer ranks, and ideally a sweep over tau.","section":"Section III, Algorithm 1"},{"comment":"The experimental protocol is under-specified. The text does not state how convolutional kernels are reshaped into matrices for SVD, which layers are compressed (all layers or only conv/FC layers), whether batch-normalization statistics or biases are adjusted after compression, whether any fine-tuning is applied, or how inference time is measured (hardware, batch size, number of runs). In addition, Tables II and III report no error bars or repeated runs, so the small F1 differences (e.g., 80.1 vs 80.4 for ResNet18 on CIFAR-10) cannot be assessed for statistical significance. These omissions block verification of the claimed accuracy and speed advantages.","section":"Section IV"}],"minor_comments":[{"comment":"Table I and the surrounding text attribute accuracy drops and complexity estimates to existing methods without citations in the table, and the references cited for these methods are not always appropriate (e.g., [10] is a Fast R-CNN object-detection paper used for TSVD). The authors should cite the actual low-rank network-compression literature (e.g., Denton et al., Jaderberg et al., Zhang et al.) and provide a fair description of baseline compression ratios and accuracies from those papers.","section":"Section II-B"},{"comment":"The definition p_i = s_i / sum_j s_j assumes that not all singular values are zero; the paper should state how zero singular values are handled in the entropy calculation.","section":"Section III"},{"comment":"The first paragraph says 'The large number of hyperparameters in DNNs' where 'parameters' is meant; this should be corrected.","section":"Introduction"},{"comment":"The figure captions are generic and do not state the metric shown, the axes, the units, or the normalization; for Figure 2, the hardware, batch size, and number of trials used for timing should be reported.","section":"Figures 1 and 2"}],"recommendation":"reject","confidential_remarks":"The manuscript is written at the level of an extended abstract: the central comparison against fixed-rank SVD is missing, the key threshold tau is undisclosed, and no code or detailed protocol is provided. These are load-bearing omissions for the paper's main claim, and they cannot be repaired by minor edits."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nThe paper's core idea is simple: instead of picking a fixed rank for every layer, choose the smallest k such that the cumulative spectral entropy reaches a fraction tau of the total. That is a legitimate, if small, variant of the standard energy-retention rule. The framing is clean, the pseudocode is unambiguous, and the complexity analysis (O(k(m+n)) per layer) is correct. So there is something here worth one careful look.\n\nThe problem is that the experiments, as written, do not test the claim. In Table II the rows labeled 'SVD' for ResNet18, InceptionV3, and VGG16 are 11.2M, 23.9M, and 134.3M parameters. Those are exactly the parameter counts of the uncompressed pretrained models. That means the authors compared ARSVD against full models, not against a fixed-rank SVD truncation. The abstract and introduction claim ARSVD beats 'traditional fixed-rank SVD,' but no fixed-rank experiment appears anywhere. This is not a stylistic weakness; it is the entire empirical case. On top of that, the entropy threshold tau is never disclosed, the reshaping of convolutional kernels into matrices is unspecified, and there are no error bars, repetition counts, or code. The F1 improvements (e.g., 80.1 to 80.4 on CIFAR-10) are within what noise alone could produce, and without a compressed baseline there is no way to know whether the method preserves accuracy better than any rank-selection rule.\n\nThe theoretical claim that entropy fraction tracks task-relevant information is asserted rather than derived, but that is a minor issue compared with the missing baseline. The definitional circularity the reader flags is real but mild: of course the method retains tau of the spectral entropy by construction; what matters is whether that correlates with performance, which the experiments could have shown and currently do not.\n\nBottom line: the paper is a clear, readable write-up of a reasonable incremental idea, and the authors should be encouraged to fix the baseline, disclose tau and the reshaping, and release code. As it stands, the central claim is untested. I would not cite it in current form, but I would send it to a serious referee rather than desk-reject, because the underlying method is well-defined and a corrected version could be a small but legitimate contribution.","headline":"Entropy-based adaptive rank selection is a real but incremental idea; the paper's central comparison to fixed-rank SVD is missing because the 'SVD' baseline rows are the uncompressed models.","tokens_in":7620,"tokens_out":1866,"would_cite":false,"duration_ms":17952,"reading_group":"no","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 spectral entropy should set each layer's SVD truncation rank, and reports that this beats fixed-rank SVD compression on pretrained CNNs.","keywords":["neural network compression","low-rank approximation","singular value decomposition","adaptive rank selection","spectral entropy","model efficiency","deep learning"],"falsifier":"Take one pretrained ResNet18, apply ARSVD with the paper's threshold, and for each resulting total parameter count apply fixed-rank SVD with the uniform rank that hits the same budget; plot F1 against parameters. If ARSVD sits below the fixed-rank curve at equal parameter budgets, the central claim is false; if the only comparison available is against full-rank models, the claim remains untested.","tokens_in":6592,"feed_emoji":"📉","tokens_out":12566,"duration_ms":118743,"temperature":0.7,"pith_summary":"ARSVD compresses a pretrained neural network by giving each layer its own SVD rank instead of using one fixed rank everywhere. For every weight matrix, it normalizes the singular values into a probability distribution, computes the spectral entropy of the full spectrum, and keeps the smallest number of top singular values whose partial entropy reaches a fraction of the total. The paper claims this layer-wise choice removes redundancy while preserving each layer's information, and reports that it beats fixed-rank SVD on ResNet18, VGG16, and InceptionV3 across MNIST, CIFAR-10, and CIFAR-100. In the reported experiments, parameter counts fall by roughly 1.5 to 28 million depending on the model, with F1 scores essentially unchanged or slightly higher. The payoff would be cheaper storage and faster inference for already-trained models without fine-tuning.","feed_headline":"Spectral entropy sets each layer's SVD rank, trimming pretrained networks","feed_subtitle":"One spectral-entropy threshold picks each weight matrix's rank, cutting millions of parameters without lowering reported F1.","key_machinery":"The engine of the method is the spectral entropy of a weight matrix's singular values. For $W\\in\\mathbb{R}^{m\\times n}$ with singular values $s_1\\ge\\cdots\\ge s_r$, the paper forms normalized weights $p_i = s_i/\\sum_j s_j$, computes total entropy $H_{\\mathrm{total}} = -\\sum_i p_i \\log p_i$, and computes the partial entropy $H(k)$ of the top $k$ singular values. The selected rank is the smallest $k$ with $H(k)\\ge \\tau H_{\\mathrm{total}}$. This object does the work of mapping a matrix to a rank without labels or gradients: a peaked spectrum triggers aggressive truncation, while a flat spectrum triggers mild truncation.","core_discovery":"On its own terms, the paper's discovery is that the singular value spectrum of a layer carries enough information to choose its compression rank. Defining $p_i = s_i/\\sum_j s_j$ and $H(k) = -\\sum_{i=1}^k p_i \\log p_i$, ARSVD picks the smallest $k$ with $H(k) \\ge \\tau H_{\\mathrm{total}}$ and replaces $W = USV^T$ by $\\tilde W = U_k S_k V_k^T$. The paper argues that this retains at least a fraction $\\tau$ of the spectral information of every layer, so layers with concentrated spectra are compressed hard and layers with flat spectra are barely compressed. It reports that this adaptive truncation lowers storage from $\\sum_l m_l n_l$ to $\\sum_l k_l(m_l+n_l)$, lowers the forward pass from $O(mn)$ to $O(k(m+n))$ per layer, and produces F1 scores equal to or slightly above the SVD baselines it compares against.","pith_inferences":["A direct test of the entropy premise would apply ARSVD to transformer weight matrices in BERT or ViT and compare with fixed-rank SVD at equal parameter budgets; the paper names transformers as future work but does not run this test.","Because the selected rank rises monotonically as $\\tau$ approaches 1, $\\tau$ could be binary-searched to hit a target parameter count, turning ARSVD into a budget-constrained compressor rather than an accuracy-first one.","A held-out validation search over per-layer ranks would separate what spectral entropy contributes from what any parameter reduction contributes; if grid-searched ranks beat entropy-chosen ranks at the same size, entropy alone is not selecting the task-relevant information.","The reported 'SVD' parameter counts match the full uncompressed models' parameter counts, so a head-to-head against genuinely truncated fixed-rank SVD is the experiment needed to support the paper's comparative claim."],"forward_implications":["Every pretrained layer can be compressed immediately, with no fine-tuning, by computing its SVD once and keeping the top $k$ singular values selected by the entropy rule.","Model storage drops from $\\sum_l m_l n_l$ to $\\sum_l k_l(m_l+n_l)$, and the forward pass cost per layer drops from $O(mn)$ to $O(k(m+n))$, so the largest absolute savings appear in wide layers such as VGG16's fully connected blocks.","Layers with highly concentrated spectra are compressed aggressively, while layers with flat spectra keep more singular values, so the same global threshold produces different compression ratios across an architecture.","Because the rank choice uses only the weight matrix, the method requires no training labels, gradients, or loss information, making it a drop-in step after any pretraining pipeline."],"supporting_citations":[{"why":"Survey that motivates low-rank matrix approximation as a standard route to neural-network compression.","marker":"[3]"},{"why":"Truncated SVD baseline whose fixed-rank truncation ARSVD is intended to outperform.","marker":"[10]"},{"why":"Randomized SVD baseline used in the complexity and accuracy comparison.","marker":"[11]"},{"why":"ℓp-norm approximation that serves as the accuracy-preserving baseline in the comparison table.","marker":"[13]"},{"why":"Adaptive pruning-rate method that motivates per-layer, sensitivity-guided compression decisions.","marker":"[18]"}],"fun_headline_variants":["Adaptive-rank SVD from spectral entropy slashes network size","Entropy-guided SVD assigns each layer its ideal compression rank","Layer-wise SVD ranks from spectral entropy cut network parameters","Singular value entropy picks per-layer ranks, trimming DNNs efficiently","Spectral entropy sets per-layer SVD ranks, trimming pretrained nets"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The claim depends on two premises the paper does not demonstrate: one entropy threshold $\\tau$ preserves the classification-relevant information of every reshaped layer, and the 'SVD' baseline is a genuinely truncated model rather than the full pretrained network.","fun_headline_variants_meta":{"raw":{"variants":["Adaptive-rank SVD from spectral entropy slashes network size","Entropy-guided SVD assigns each layer its ideal compression rank","Layer-wise SVD ranks from spectral entropy cut network parameters","Singular value entropy picks per-layer ranks, trimming DNNs efficiently","Spectral entropy sets per-layer SVD ranks, trimming pretrained nets"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000904,"raw_usage":{"total_tokens":3852,"prompt_tokens":871,"completion_tokens":2981,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":487,"completion_tokens_details":{"reasoning_tokens":2892}},"tokens_in":487,"tokens_out":2981,"duration_ms":21628,"temperature":1.0,"reasoning_tokens":2892,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-16T10:24:05.314148+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take one pretrained ResNet18, apply ARSVD with the paper's threshold, and for each resulting total parameter count apply fixed-rank SVD with the uniform rank that hits the same budget; plot F1 against parameters. If ARSVD sits below the fixed-rank curve at equal parameter budgets, the central claim is false; if the only comparison available is against full-rank models, the claim remains untested.","supporting_citations":[{"cited_title":"The power of convex relaxation: Near-optimal matrix completion,","cited_arxiv_id":null,"evidence_quote":"ℓp-norm approximation that serves as the accuracy-preserving baseline in the comparison table."},{"cited_title":"Structured pruning for deep neural networks with adaptive pruning rate derivation based on connection sensitivity and loss function,","cited_arxiv_id":null,"evidence_quote":"Adaptive pruning-rate method that motivates per-layer, sensitivity-guided compression decisions."}],"review_version":1}