{"id":"a54d17c3-cc53-4255-8f43-8540fb25a28c","arxiv_id":"2507.22527","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"FGFP combines seven-parameter fractional Gaussian filters with adaptive unstructured pruning to compress CNNs by 69-97% with only about 1-2% accuracy loss.","lead":"This paper introduces FGFP, a method that replaces convolutional kernels with fractional Gaussian filters using only seven parameters per kernel, then prunes the remaining weights. The authors report model-size reductions of 69-97% on CIFAR-10 and ImageNet with accuracy drops of about 1-2%.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The paper never specifies how pretrained kernels are converted to seven-parameter FGFs; without that procedure the reported 85.2%/69.1% compression results cannot be reproduced or attributed to the proposed representation.","rationale":"After reading the manuscript, the weakest point is precisely what the reader identified. The paper's novelty and results rest on replacing kernels with seven-parameter FGFs, but the replacement procedure is not described, and the mathematical approximation that justifies the FGF form is asserted without error control. Without this, any accuracy/compression number could depend on an undocumented conversion or fine-tuning recipe, so the central claim is not yet substantiated. I do not see a more basic flaw: the FGF factorization is algebraically valid as a parameterization, and the reported numbers are not internally impossible. The baseline mismatches (e.g., WRN-28-10 baselines differ by 1.42 accuracy points) and unreported pruning thresholds are real but secondary; they affect the 'outperforms recent methods' subclaim, not the core feasibility of the compression. A conditional acceptance with a request for code and the conversion algorithm is therefore appropriate. If the code reproduces the table and per-layer fit error is small, the concern is resolved; if not, the central claim is unsupported.","tokens_in":10524,"tokens_out":7381,"duration_ms":94070,"concrete_test":"Request the released code or a precise pseudocode for Section 2.4. Then, on ResNet-20/CIFAR-10, run the full FGFP pipeline at the Table 1 setting (CR 85.2, 3D-FGF), recording two diagnostics: (i) per-layer relative Frobenius error between the original kernels and the best seven-parameter FGF fit before fine-tuning, and (ii) final compressed accuracy. If the final accuracy drop is not within the reported 1.52% at 85.2% CR, or if the initial FGF fit error is large (say median >10%), then the missing conversion step is load-bearing for the central claim.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim—85.2% parameter CR with 1.52% accuracy drop on ResNet-20 and 69.1% CR with 1.63% drop on ResNet-50—requires that a pre-trained convolution kernel can be replaced by the seven-parameter FGF form in Eq. (11) without incurring the accuracy loss that the restrictive outer-product structure would naively imply. The paper never supplies the step that would establish this. Section 2.4 says only 'we convert the filters of the pre-trained model into the FGF' and Fig. 1 shows a 'FGF Conversion Process', but the actual algorithm is absent: there is no objective function, no initialization scheme, and no description of whether FGF parameters are fitted to the original kernels or randomly initialized and then trained. Additionally, the truncation in Eq. (4) drops the error term with the sentence 'the error can be ignored' and offers no bound. Because the central empirical numbers depend on this unspecified conversion and on the dropped approximation, the headline compression/accuracy trade-off is not independently checkable from the manuscript. This is a missing-support concern, not a demonstration that the method fails.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"This paper introduces FGFP, a compression framework that replaces selected convolutional kernels with fractional Gaussian filters formed as products of truncated Grünwald–Letnikov fractional derivatives of Gaussian functions (seven parameters per 3D-FGF kernel, Eq. 11), and then applies adaptive unstructured pruning to the remaining layers. The manuscript reports strong empirical trade-offs: on CIFAR-10, ResNet-20 achieves a 1.52% accuracy drop with 85.2% parameter compression, and on ImageNet2012, ResNet-50 achieves a 1.63% drop with 69.1% compression. The authors claim these results outperform recent pruning and low-rank compression methods.","tokens_in":10796,"tokens_out":7336,"duration_ms":86228,"significance":"If the reported trade-offs are reproducible, the method is a novel and practically useful compression scheme for convolutional networks: the seven-parameter 3D-FGF parameterization is an elegant way to obtain very high parameter compression, and the ablation in Table 3 shows that combining FGF with AUP outperforms AUP alone at the same parameter budget. The paper also gives careful parameter-count accounting and reports experiments on both CIFAR-10 and ImageNet2012. However, the missing conversion algorithm and the unanalyzed truncation error currently prevent independent verification, and some baseline comparisons are not matched; these issues are central rather than cosmetic.","major_comments":[{"comment":"The paper never specifies how a pretrained convolutional kernel is mapped to the seven-parameter FGF form of Eq. (11). Section 2.4 states only that \"we convert the filters of the pre-trained model into the FGF\" and Figure 1 labels a \"FGF Conversion Process\", but no algorithm is given: there is no objective function, initialization scheme, optimizer, or statement of whether the FGF parameters are fitted to the original weights or are randomly initialized and then trained end-to-end. Because the headline compression/accuracy results rest on this conversion, the method is not reproducible from the manuscript in its current form.","section":"Section 2.4 and Figure 2"},{"comment":"The truncation of the Grünwald–Letnikov series is not justified. Equation (4) drops the error term epsilon with the sentence \"the error can be ignored\", but no bound, numerical estimate, or ablation is supplied. Since all FGF forms in Eqs. (9)-(11) are built from the three-term approximation, the validity of the fitted FGF representation depends on this dropped term; please quantify the approximation error over the allowed ranges a,b,c ∈ [0,2] or provide an ablation with additional series terms.","section":"Equations (4)-(5) and (9)-(11)"},{"comment":"The WRN-28-10 comparison is not apples-to-apples. GrowEfficient and BackSparse report baselines of 96.20% (with deltas marked as validation accuracy), while the FGFP WRN-28-10 baseline is 94.78%, so a smaller accuracy drop from a lower baseline does not establish superiority. Similar baseline discrepancies appear in Table 2 for ResNet-50 (baselines of 75.90, 76.15, and 76.20 across methods). Please retrain or re-evaluate under a matched protocol and clearly separate validation-based numbers from test accuracy.","section":"Table 1 (WRN-28-10) and Table 2"},{"comment":"The empirical claims are reported as single points without error bars, number of runs, or seeds. In addition, Section 3.1 does not specify the fine-tuning schedule for the pruning rounds, the rule for reducing pr, the number of rounds, or how the \"deeper layers with larger input channels\" are selected for conversion. These details are needed to assess whether the reported accuracy/compression trade-offs are stable and not due to favorable initialization or dataset-specific settings.","section":"Section 3.1 and Tables 1-4"}],"minor_comments":[{"comment":"The headline numbers are inconsistent: the abstract reports a 1.52% drop with 85.2% compression for ResNet-20, while Section 1 reports a 1.58% drop with 85.1% compression; Table 1 agrees with the abstract, so the introduction should be corrected.","section":"Abstract vs. Introduction"},{"comment":"The figure caption and surrounding text refer to \"FGFS\" in several places (e.g., \"Overview of the FGFS methodology\", \"the FGFS contains two FGF forms\"), while the method is named FGFP elsewhere; please unify the terminology.","section":"Figure 1 and Section 2.4"},{"comment":"There is a typo in the spelling of Grünwald in Equation (4) (\"Grüwald\"), and the same equation uses epsilon with an unusual subscript placement; please clean up the notation.","section":"Equation (4)"},{"comment":"The AUP algorithm resembles iterative magnitude pruning but is not compared with standard magnitude pruning or lottery-ticket-style rewinding; such a comparison would help isolate the contribution of the FGF parameterization from the pruning strategy.","section":"Section 2.3"},{"comment":"No code, checkpoints, or configuration files are referenced, which makes it difficult to verify the reported results independently of the textual description.","section":"Reproducibility"}],"recommendation":"major_revision","confidential_remarks":"The core idea is interesting and the reported numbers are internally coherent, but the missing weight-to-FGF conversion algorithm is a reproducibility blocker and the WRN-28-10 comparison is not matched. I would be willing to see a revised version that adds pseudocode or an appendix for the conversion, quantifies the GL truncation error, and reruns the WRN-28-10 experiments under the same baseline protocol."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nThe honest summary: there is a real idea here, and the paper as written does not let anyone check it. The 3D-FGF construction — seven parameters per kernel, with a fractional derivative of a Gaussian along the channel dimension — is a clean extension of Zamora et al.'s fractional Gaussian filter, and pairing it with adaptive unstructured pruning produces a genuinely new compressed-architecture family. The ablation where FGFP beats AUP alone at equal parameter count is the right kind of evidence that the filter part contributes.\n\nThe math is elementary but the formulation is coherent. The Grünwald-Letnikov truncation to three terms is standard; the dropped error term in Eq. (4) is not bounded, but this looks like a minor risk in practice, not a design flaw.\n\nThe load-bearing gap is the conversion procedure. Section 2.4 says \"we convert the filters of the pre-trained model into the FGF\" and points to the diagram, but never gives the algorithm: no objective function, no initialization, no description of whether the FGF parameters are fitted to the original kernels or randomly initialized and trained. Since the headline numbers (85.2% CR with 1.52% drop on ResNet-20; 69.1% with 1.63% on ResNet-50) depend entirely on this step, the results are not reproducible from the manuscript. This is a missing-support problem, not evidence that the method fails.\n\nThe other weaknesses are smaller. The WRN-28-10 comparison in Table 1 uses a different post-trained baseline (94.78 vs 96.20 for the compared works), so the reported accuracy-drop advantage is partly an artifact of the baseline choice. There are no error bars, and a couple of numbers don't match between the abstract and the contributions paragraph (1.52/85.2 vs 1.58/85.1). These are fixable.\n\nIf the authors ship code and the conversion details, I would take the claims seriously. As it stands, treat this as a promising work-in-progress, not a verified result. Worth a reading group conversation about what a compression paper owes its readers. A serious referee should ask for the conversion algorithm and code before the claims can be weighed; that is a normal revision request, not a rejection.\n\nBest,","headline":"A promising compression idea whose central conversion step is missing, so the headline numbers are unverified as written.","tokens_in":11339,"tokens_out":3320,"would_cite":false,"duration_ms":34614,"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":"FGFP compresses ResNet-20 to 14.8% of its original size with a 1.52% accuracy drop by replacing full kernels with seven-parameter fractional Gaussian filters and pruning the rest.","keywords":["network compression","fractional Gaussian filter","Grünwald–Letnikov derivative","unstructured pruning","ResNet","CIFAR-10","ImageNet2012","model size reduction"],"falsifier":"Take a pretrained ResNet-20, convert its chosen layers to 3D-FGF parameters, reconstruct the explicit kernels from Eq. (11), and measure the mean relative Frobenius-norm error against the original weights; if the error is large, or if evaluating the converted model before any pruning shows an accuracy drop far above the reported 1.52%, the claim that FGFs carry the compression would be falsified.","tokens_in":10293,"feed_emoji":"📉","tokens_out":7552,"duration_ms":83871,"temperature":0.7,"pith_summary":"The paper proposes a compression framework that replaces full convolution kernel weights with fractional Gaussian filters (FGFs), kernels built as products of shifted Gaussian functions acted on by fractional derivatives, so each kernel costs only seven learned parameters, plus a layer-wise magnitude-pruning pass that removes small weights from the remaining layers. The claim is that this combination shrinks networks substantially while keeping accuracy nearly intact: on CIFAR-10, ResNet-20 loses 1.52 percentage points while cutting model size by 85.2%, and on ImageNet2012, ResNet-50 loses 1.63 points while cutting model size by 69.1%. If the claim holds, a sizeable part of a trained network's parameters are redundant in a specific sense: many filters can be summarized by a smooth, separable, Gaussian-derived formula. That would make deep models considerably cheaper to store and to load on memory-constrained devices.","feed_headline":"A seven-parameter filter compresses ResNets by up to 85%","feed_subtitle":"Fractional Gaussian filters plus sparse pruning cut ResNet-50 to 31% of its size with a 1.63% accuracy drop.","key_machinery":"The load-bearing object is the three-dimensional fractional Gaussian filter (3D-FGF): a rank-1 separable kernel expressed as the product of Grünwald–Letnikov fractional derivatives of 1D Gaussians along the $x$, $y$, and channel axes. Its seven parameters replace an entire $n \\times n \\times ch$ weight tensor, so the parameter count collapses from $9\\cdot ch$ (for a $3\\times 3$ filter) to 7. The channel-attention variant (CA-FGF) instead keeps five shared filter parameters plus one learned weight per channel, trading a little compression for more accuracy. Adaptive unstructured pruning (AUP) then removes small-magnitude weights in rounds, with fine-tuning after each round and rollback if the accuracy threshold is not met. Together the two mechanisms are what the experiments claim outperform low-rank decomposition and pruning baselines.","core_discovery":"The central discovery, stated on the authors' terms, is that a convolution filter can be re-parameterized as the product of three Grünwald–Letnikov fractional derivatives of one-dimensional Gaussian functions, $F_{3d} = D_x^a G(x) \\times D_y^b G(y) \\times D_{ch}^c G(ch)$, with each fractional derivative approximated by a trinomial $D^\\alpha f(x) \\approx f(x) - \\alpha f(x-1) + \\frac{\\alpha(\\alpha-1)}{2} f(x-2)$. Sharing these parameters across all input channels and along the channel direction leaves only seven trainable numbers per kernel: three derivative orders $a,b,c$, two spatial centers plus a channel center, and the Gaussian width $\\sigma$. The paper further claims that adding this FGF conversion to a simple adaptive unstructured pruning loop, which prunes low-magnitude weights, fine-tunes, and rolls back if accuracy drops, yields accuracy equal to or better than recent low-rank and pruning-only methods at the same or higher compression ratios.","pith_inferences":["Inference: the FGF's functional form is a strong inductive prior, saying useful learned filters are smooth, separable, and close to Gaussian-derivative shapes, so the method will likely compress best in layers whose kernels resemble classical image filters and worst in layers encoding sharp, non-separable features.","Inference: a direct reconstruction test would separate the two contributors to the reported numbers: how accurately the seven-parameter form reproduces the original kernels, versus how much accuracy is recovered by fine-tuning and by AUP; the paper does not isolate these.","Inference: the framework could be composed with structured pruning, quantization, or distillation to push memory reduction further, since FGF handles filter-level redundancy while AUP handles element-level redundancy.","Inference: parameter count is not the same as inference cost; a natural follow-up is to measure latency and energy on mobile CPUs and GPUs, where the separable Grünwald–Letnikov approximation may or may not translate into speedups."],"forward_implications":["A ResNet-50 on ImageNet2012 would shrink from about 25.6 million parameters to about 7.9 million, making it feasible to store the compressed model in on-device memory.","Training the compressed model is cheap because entire kernels are governed by seven numbers, so the FGF layers have very few learnable parameters before pruning.","The ablation on ResNet-20 suggests FGF conversion adds value beyond pruning alone: at 0.07M remaining parameters, FGFP with CA-FGF loses 0.86 accuracy points versus 1.19 for AUP by itself.","The same recipe applies across architectures and scales, since the paper reports results for ResNet-20/32, WRN-28-10, ResNet-18, and ResNet-50.","Because the remaining layers are unstructured-sparse, the final model can be stored with sparse formats, which is what turns the parameter reduction into actual memory savings on edge hardware."],"supporting_citations":[{"why":"Introduces fractional derivatives of Gaussian filters inside CNNs, the starting point for the FGF construction.","marker":"Zamora et al., 2021"},{"why":"Supplies the Grünwald–Letnikov fractional derivative variant and the trinomial approximation that reduces FGF computation.","marker":"Jalalinejad et al., 2018"},{"why":"Provides the Grünwald–Letnikov method for fractional differential equations, the mathematical basis for Eq. (4).","marker":"Scherer et al., 2011"},{"why":"Defines the ResNet-20/32/18/50 architectures used for all main experiments and baselines.","marker":"He et al., 2016"},{"why":"PSTRN-M is the CIFAR-10 low-rank baseline at the roughly 85% compression ratio that FGFP must beat.","marker":"Li et al., 2022"},{"why":"FR is the ImageNet ResNet-18 low-rank baseline that FGFP compares against at 60–75% compression.","marker":"Chu & Lee, 2021"},{"why":"Supports the layer-selection strategy that FGF conversion is most effective in deeper layers with larger input channels.","marker":"Llanza et al., 2023"}],"fun_headline_variants":["Fractional Gaussian filter: 7 params, 85% smaller models","Seven-parameter filter plus pruning cuts ResNet size 85%","A 7-parameter kernel shrinks ResNet-20 by 85.2%","FGFP: 7-parameter filter, 85% compression, 1.5% drop","Fractional Gaussian filter and pruning: 85% size reduction"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The whole compression rests on the premise that real pretrained convolution kernels can be accurately written as a product of three truncated fractional derivatives of Gaussians, seven numbers per kernel, with the omitted tail of the Grünwald–Letnikov approximation negligible; the paper does not directly verify this representation.","fun_headline_variants_meta":{"raw":{"variants":["Fractional Gaussian filter: 7 params, 85% smaller models","Seven-parameter filter plus pruning cuts ResNet size 85%","A 7-parameter kernel shrinks ResNet-20 by 85.2%","FGFP: 7-parameter filter, 85% compression, 1.5% drop","Fractional Gaussian filter and pruning: 85% size reduction"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000802,"raw_usage":{"total_tokens":3553,"prompt_tokens":1000,"completion_tokens":2553,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":616,"completion_tokens_details":{"reasoning_tokens":2449}},"tokens_in":616,"tokens_out":2553,"duration_ms":20599,"temperature":1.0,"reasoning_tokens":2449,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T11:34:33.212141+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take a pretrained ResNet-20, convert its chosen layers to 3D-FGF parameters, reconstruct the explicit kernels from Eq. (11), and measure the mean relative Frobenius-norm error against the original weights; if the error is large, or if evaluating the converted model before any pruning shows an accuracy drop far above the reported 1.52%, the claim that FGFs carry the compression would be falsified.","supporting_citations":[{"cited_title":"A., Rhodes, A., Nachman, L., and Sundararajan, N","cited_arxiv_id":null,"evidence_quote":"Introduces fractional derivatives of Gaussian filters inside CNNs, the starting point for the FGF construction."},{"cited_title":"A simple and flexible modification of gr \\\"u nwald--letnikov fractional derivative in image processing","cited_arxiv_id":null,"evidence_quote":"Supplies the Grünwald–Letnikov fractional derivative variant and the trinomial approximation that reduces FGF computation."},{"cited_title":"L., Tang, Y., and Huang, J","cited_arxiv_id":null,"evidence_quote":"Provides the Grünwald–Letnikov method for fractional differential equations, the mathematical basis for Eq. (4)."},{"cited_title":"Low-rank Tensor Decomposition for Compression of Convolutional Neural Networks Using Funnel Regularization","cited_arxiv_id":"2112.03690","evidence_quote":"FR is the ImageNet ResNet-18 low-rank baseline that FGFP compares against at 60–75% compression."},{"cited_title":"E., Shvai, N., and Nakib, A","cited_arxiv_id":null,"evidence_quote":"Supports the layer-selection strategy that FGF conversion is most effective in deeper layers with larger input channels."}],"review_version":1}