{"id":"3bc14bc1-10b2-4550-ae34-1722e87aaf0a","arxiv_id":"2411.18388","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"A ResNet variant with frozen edge filters and learned 1x1 combinations matches or beats ResNet18 on several fine-grained datasets while using 13% of its parameters.","lead":"This paper shows that a convolutional network can recognize images even when its spatial filters are fixed in advance and never learned, with only the way their outputs are combined trained. The result suggests much of a CNN's power is in learning feature combinations, not in the filters themselves, and could lead to smaller models.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Headline 'PFNet18 outperforms ResNet18' rests on an uneven comparison: PFNet18 receives an anti-aliasing blur that ResNet18 does not, so the claimed evidence for pre-defined filters is overstated.","rationale":"The reader's formal weakest_assumption concerns the sufficiency of the 16 hand-picked edge filters, but the reader's rationale already identifies the inconsistent anti-aliasing treatment as a primary concern. I agree that the anti-aliasing issue is the most concrete and load-bearing weakness in the empirical argument: the strongest claim that pre-defined spatial filters can match or beat ResNet18 is directly supported by Table 4, and that table appears to compare a no-aliasing PFNet18 against a default ResNet18. The papers own Table 7 provides the crucial missing data point for Flowers102 (ResNet18 no-aliasing = 74.36), but no such data are given for the other two datasets where PFNet18 is reported to win by large margins. This is a methodological asymmetry, not a flaw in the central mechanism: the anti-aliasing blur is pre-defined and frozen, so it does not violate the 'only 1x1 combinations are learned' restriction. The central claim that pre-defined filters can work remains plausible, and the paper includes useful ablations (random frozen filters, trainable filters) that partially de-risk the filter-choice assumption. However, the headline empirical comparison should be rerun under identical conditions before the superiority claim is accepted. This concern is addressable by a targeted experiment, so a conditional acceptance is appropriate; hence I leave the reader's CONDITIONAL verdict unchanged. My disagreement with the formal weakest_assumption field is partial because I view the anti-aliasing asymmetry, not the filter-set sufficiency, as the most load-bearing issue; the ablation study already provides some evidence on filter-set sensitivity, whereas the anti-aliasing asymmetry directly affects the papers central comparative evidence.","tokens_in":10968,"tokens_out":12880,"duration_ms":112894,"concrete_test":"Retrain ResNet18 with the same 3x3 Gaussian blur applied to the three stride-2 1x1 convolution inputs, as described in Section 5.4 for PFNet18, on Caltech101, FGVC-Aircraft, and Flowers102, using the same optimizer, epochs, and augmentation as Table 4. Compare the resulting accuracies with PFNet18's Table 4 entries. If the no-aliasing ResNet18 matches or exceeds PFNet18 on any dataset, the 'outperforms' claim is unsupported for that dataset; if PFNet18 still wins on all three, the anti-aliasing asymmetry is not the source of the reported gains and the headline comparison is robust.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central evidence that pre-defined filters suffice is the claim that PFNet18 matches or beats ResNet18 (Table 4). Section 5.4 and Table 7 reveal that PFNet18's Flowers102 result of 80.66 is the 'no aliasing' configuration, obtained by blurring the inputs to the three stride-2 1x1 convolution operations with a 3x3 Gaussian filter. The ResNet18 result in Table 4 (73.40) is the default configuration without that blur; the paper's own no-aliasing ResNet18 is 74.36. For Caltech101 and FGVC-Aircraft, where PFNet18 is reported to outperform ResNet18 by large margins, no no-aliasing ResNet18 results are reported. If the default PFNet18 used for those datasets also includes the anti-aliasing blur, then the comparison is asymmetric: PFNet18 gets a pre-defined preprocessing step that ResNet18 does not get. This does not refute the central idea that spatial kernels need not be learned, because the blur is itself a pre-defined, frozen operation, and the no-aliasing ResNet18 on Flowers still trails PFNet18. However, the headline superiority claim is currently supported only under unequal conditions, and the magnitude of the reported gains on Caltech101 and FGVC-Aircraft is unverified under fair comparison.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper introduces Pre-defined Filter Convolutional Neural Networks (PFCNNs), in which every spatial convolution kernel of size greater than 1 is drawn from a fixed pool of 16 hand-selected 3x3 edge filters and remains constant during training; only 1x1 convolutions are learned. The authors build PFNet18 by replacing all n>1 convolutions of ResNet18 with Pre-defined Filter Modules (PFMs), reducing the parameter count to 1.46M (13% of ResNet18). They report five-seed experiments on Caltech101, CIFAR10, CUB-200-2011, FGVC-Aircraft, Flowers102, and Stanford Cars, together with feature visualizations, an ablation on Flowers102, and an analysis of aliasing. The central empirical claim is that PFNet18 matches or outperforms ResNet18 on several datasets despite the severe restriction on learnable spatial kernels.","tokens_in":11244,"tokens_out":8084,"duration_ms":65159,"significance":"If the results hold under a fair comparison, the paper makes a compelling empirical case that deep networks can achieve competitive recognition accuracy by learning only linear combinations of a small number of pre-defined spatial filters, which is a valuable contribution to the debate about what CNNs actually learn. The study is transparent: code is released, results are averaged over five seeds, and the authors include an ablation separating the effect of filter choice from the fixed-filter mechanism. The main reservation is that the headline superiority of PFNet18 is currently supported by an unequal comparison in which PFNet18 receives a Gaussian anti-aliasing blur that ResNet18 does not receive, and the reported 10% improvement exceeds the largest margin in the data; the significance is therefore conditional on correcting these issues.","major_comments":[{"comment":"Table 4 reports PFNet18 Flowers102 accuracy of 80.66, which Table 7 identifies as the \"No aliasing, default\" configuration, while the ResNet18 value of 73.40 is the default (aliasing) configuration. For Caltech101 and FGVC-Aircraft, the manuscript does not state whether the PFNet18 results use the anti-aliasing blur, and no no-aliasing ResNet18 baselines are reported for these datasets. The claimed superiority of PFNet18 on these datasets is therefore not established under equal conditions. Please either report ResNet18 no-aliasing results for all datasets in Table 4, or use the aliasing version of PFNet18 for the main comparison, and state explicitly which configuration was used for every entry.","section":"Section 5.4, Tables 4 and 7"},{"comment":"The abstract and Section 7 state that PFNet18 outperforms ResNet18 \"with an absolute increase of 10%\". The largest gap in Table 4 is 8.41 percentage points (Caltech101: 65.60 vs 57.19); Flowers102 is 7.26 points and FGVC-Aircraft is 2.17 points. The claim should be reworded to report the actual margins, and it should be checked again after the aliasing treatment is aligned across models.","section":"Abstract and Section 7"},{"comment":"The claim that \"ResNet18 has a lower risk of losing image information by aliasing\" and that PFNet18 relies on skip connections to compensate is supported only by a single quantitative experiment on Flowers102 plus the small illustrative example in Figure 3. If the anti-aliasing blur is intended to be part of the default PFNet18 configuration, its effect on Caltech101, FGVC-Aircraft, and the other datasets should be quantified so that the reader can separate the benefit of the fixed filters from the benefit of the blur.","section":"Section 5.4"}],"minor_comments":[{"comment":"The permutation argument for f=1 is only sketched; please spell out how a permutation of kernel assignments is absorbed by the weights of the preceding and following 1x1 layers, especially in the presence of residual connections.","section":"Section 3.2"},{"comment":"The CIFAR10 paragraph states that \"these adjustments apply for both, PFNet18 and ResNet18\", which is ambiguous because PFNet18's first convolution is already 1x3x3; please specify the exact architectural modifications for each model.","section":"Section 4.1"},{"comment":"The filter-choice ablation is conducted only on Flowers102; the general conclusion in Section 6 that edge filters add a suitable bias would be strengthened by repeating the random-versus-edge comparison on at least one additional dataset.","section":"Section 5.5"},{"comment":"The weight decay value is reported as 1.0; please confirm this is intentional and not a typo for 1e-4, since it is an unusual setting.","section":"Section 4.2"},{"comment":"Figure 3 is difficult to read at print size; please enlarge the panels and add explicit labels for the filters and the input.","section":"Figure 3"},{"comment":"The caption's phrase \"During BP the hyperparameters described above are used\" is unclear; the backward-pass time is not a function of hyperparameters, so please rephrase to indicate the exact conditions of the measurement.","section":"Table 6 caption"}],"recommendation":"major_revision","confidential_remarks":"This paper appears to be a journal extension of an IJCNN 2023 paper (see the acknowledgement). The editor should verify that the new material — the full PFNet18 architecture, the additional datasets, and the aliasing analysis — constitutes sufficient novelty for this journal. The uneven anti-aliasing comparison is the main technical issue to resolve."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The paper's main empirical point is solid: you can freeze every n>1 spatial kernel to a pool of 16 hand-picked edge filters, learn only 1x1 combinations, and land close to ResNet18 on several datasets. That is not a trivial outcome, and the ablation showing edge filters beat frozen random filters (80.66 vs 73.98 on Flowers102) is the real contribution. Gavrikov and Keuper had frozen random filters; this paper shows the choice of filter matters and builds a complete ResNet18 replacement with 13% of the trainable parameters. The experiments are clean: 5 seeds, error bars, ablations for filter choice, ReLU removal, trainable vs frozen, and the aliasing discussion is unusually honest. Code is on GitHub. Credit where due: this is a worthwhile empirical study.\n\nNow the soft spots. The stress-test concern is correct and it is the paper's main weakness. Table 4 reports PFNet18 at 80.66 on Flowers102, but that number only appears after the anti-aliasing fix described in Section 5.4 — the 3x3 Gaussian blur injected before the three stride-2 skip-connection 1x1 convolutions. The ResNet18 baseline in Table 4 is the default, no-blur configuration. The no-aliasing ResNet18 on Flowers (74.36) is reported in Table 7, and PFNet18 still wins by about 6 points against that fairer baseline, so the core claim survives. But for Caltech101 and FGVC-Aircraft, where the paper claims large wins, no no-aliasing ResNet18 numbers are given, so the headline superiority is currently supported only under asymmetric conditions. This is fixable, not fatal. Second, the conclusion's \"absolute increase of 10%\" overstates the largest actual gap, which is 8.41 points on Caltech101. Third, the claim that the 8-dimensional filter space is a sufficient basis across all these datasets is only tested on Flowers; that is a minor gap, acknowledged by the future-work section.\n\nThe paper deserves a rigorous referee. The central idea is novel, the experiments are reproducible (released code), and the flaws are in the presentation of the comparison, not in the underlying conclusion. A good referee should push for symmetric no-aliasing baselines on all datasets and a more precise summary of the gains. I would cite this if I worked on parameter-efficient CNNs, and I would bring it to a reading group to talk about how easily a paper's headline can outrun its own table.\n\nRecommendation: accept to peer review, with heavy revision on the comparison details.","headline":"The core claim survives, but the headline 'PFNet18 beats ResNet18' rests on an uneven anti-aliasing comparison and a slightly inflated percentage.","tokens_in":11722,"tokens_out":2093,"would_cite":true,"duration_ms":18859,"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":"16 frozen edge filters match ResNet18 with 13% of the weights","keywords":["pre-defined filters","convolutional neural networks","depthwise convolution","edge filters","linear combinations","weight sharing","image classification","PFNet18"],"falsifier":"Train the same PFNet18 and ResNet18 from scratch on a dataset whose class-discriminative information is carried by high-frequency textures or periodic patterns rather than edges, and check whether PFNet18 falls far behind; a cleaner test is to vary the filter pool (random, Gabor, Fourier) on one dataset and show that no fixed pool restores ResNet-level accuracy on all datasets.","tokens_in":10779,"feed_emoji":"🧩","tokens_out":6837,"duration_ms":57121,"temperature":0.7,"pith_summary":"This paper tries to establish that a deep convolutional network need not learn its spatial filters to recognize images. It introduces PFCNNs, in which every spatial convolution kernel of size $n\\times n$ with $n>1$ is taken from a fixed pool of 16 edge filters and frozen; training only adjusts $1\\times1$ convolutions that linearly combine the fixed-filter outputs. On Caltech101, FGVC-Aircraft, and Flowers102, this restricted network scores higher than a standard ResNet18, and on CIFAR10, CUB-200-2011, and Stanford Cars it stays close, while using about 13% of ResNet18's trainable parameters. If the claim holds, the representational work in a CNN is largely recombination of fixed local measurements, and most trained spatial kernels are redundant.","feed_headline":"16 frozen edge filters match ResNet18 with 13% of the weights","feed_subtitle":"A network that only learns pointwise combinations of fixed 3x3 edge filters matches or beats the full model on six image sets.","key_machinery":"The Pre-defined Filter Module (PFM) carries the argument. In its channel-wise part, each input channel is convolved with one kernel from a fixed pool of 16 edge filters, assigned by the index $i \\bmod k$; an intermediate width parameter $f$ creates $f$ copies of the input channels so that the first layer applies every filter to every color channel. In its second part, a $1\\times1$ convolution learns arbitrary linear combinations of the filter outputs, so all gradient updates affect only the combination weights. The module makes the spatial filter bank shared and fixed, isolating exactly what the network has to learn.","core_discovery":"The paper's central claim is stated in its title: convolutional neural networks do work with pre-defined filters. The authors define a Pre-defined Filter Module that performs a channel-wise convolution with frozen $1\\times3\\times3$ kernels drawn from a pool of 16 hand-picked edge filters, followed by a learned $1\\times1$ convolution over all intermediate channels; replacing every $n\\times n$ convolution in ResNet18 with this module yields PFNet18. They report that PFNet18 outperforms ResNet18 on three of six benchmarks and matches it closely on the others, despite having only 1.46 million trainable parameters. The same experiments show filter choice matters: frozen random filters perform much worse than frozen edge filters, and unfreezing the edge filters gives no improvement over freezing them.","pith_inferences":["The paper leaves open whether the same 16 edge filters transfer to large-scale, non-photographic, or high-frequency-texture benchmarks; a direct test would be to train PFNet18 on ImageNet or on synthetic texture datasets where edges are not the discriminating signal.","If frozen random filters can already match ResNet18, the essential ingredient may be the learned $1\\times1$ combinations rather than the edge basis; one could test whether the edge advantage grows as training data shrinks.","The aliasing sensitivity suggests a cheap improvement: inserting an anti-aliasing blur before each stride-2 downsampling might recover most of the performance gap on datasets where PFNet18 lags, independently of any change to the filter pool.","Since the filter order is fixed and shared across channels, the architecture is a form of structured weight sharing that could be combined with model distillation or quantization; the paper does not explore this."],"forward_implications":["If spatial filters really do not need to be learned, then most of the parameters in a conventional CNN are unnecessary for image classification, and weight sharing through a small fixed filter pool can replace pruning as a compression strategy.","Because PFNet18 uses fixed edge filters in all layers, high-level object-specific features must emerge purely from linear combinations of local edge responses, which is what the feature visualizations appear to show.","Removing aliasing at stride-2 convolutions improves PFNet18 accuracy on Flowers102 from about 72% to 81%, so fixed-filter networks are more sensitive to spatial downsampling than standard CNNs.","Frozen random filters still reach ResNet18-level accuracy on Flowers102, while frozen edge filters exceed it, implying the filter basis is an inductive bias that matters most on harder small-data tasks.","The parameter reduction from 11.23M to 1.46M does not produce a comparable runtime reduction because the depthwise implementation has more graph nodes; the paper attributes this to implementation inefficiency, not to the method itself."],"supporting_citations":[{"why":"Establishes the nearest prior result that learning linear combinations of frozen random filters can work, which the paper extends to 16 hand-picked edge filters.","marker":"[8]"},{"why":"Shows pre-defined filters used in an earlier CNN, but only in the first layer, providing the comparison point for putting fixed filters throughout.","marker":"[7]"},{"why":"Defines the ResNet18 architecture that PFNet18 is built from and compared against.","marker":"[2]"},{"why":"Supplies the historical precedent of pre-defined filters in neural networks, motivating the PFCNN idea.","marker":"[6]"},{"why":"Provides empirical evidence that trained CNN filters often resemble edge filters, justifying the choice of the 16 edge kernels.","marker":"[10]"},{"why":"Offers the classic observation that first-layer filters include edge-like detectors, supporting the fixed edge-filter bias.","marker":"[9]"}],"fun_headline_variants":["16 frozen filters, 13% weights, ResNet18 accuracy","CNN with 16 pre-defined filters matches ResNet18 on 6 image sets","Only 16 edge filters, learned pointwise combos, rival full CNN","Freeze 16 filters, learn only 1x1, match full ResNet18"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that 16 hand-picked edge filters, fixed and applied in the same cyclic order in every layer, give a sufficiently rich spatial basis that no learned spatial kernel is needed on any of these datasets.","fun_headline_variants_meta":{"raw":{"variants":["16 frozen filters, 13% weights, ResNet18 accuracy","CNN with 16 pre-defined filters matches ResNet18 on 6 image sets","Only 16 edge filters, learned pointwise combos, rival full CNN","Freeze 16 filters, learn only 1x1, match full ResNet18"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001078,"raw_usage":{"total_tokens":4487,"prompt_tokens":896,"completion_tokens":3591,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":512,"completion_tokens_details":{"reasoning_tokens":3515}},"tokens_in":512,"tokens_out":3591,"duration_ms":25281,"temperature":1.0,"reasoning_tokens":3515,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T11:15:18.426754+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Train the same PFNet18 and ResNet18 from scratch on a dataset whose class-discriminative information is carried by high-frequency textures or periodic patterns rather than edges, and check whether PFNet18 falls far behind; a cleaner test is to vary the filter pool (random, Gabor, Fourier) on one dataset and show that no fixed pool restores ResNet-level accuracy on all datasets.","supporting_citations":[{"cited_title":"PCFNet: Deep neural network with predefined convolutional filters,","cited_arxiv_id":null,"evidence_quote":"Shows pre-defined filters used in an earlier CNN, but only in the first layer, providing the comparison point for putting fixed filters throughout."},{"cited_title":"Deep Residual Learning for Image Recognition,","cited_arxiv_id":null,"evidence_quote":"Defines the ResNet18 architecture that PFNet18 is built from and compared against."},{"cited_title":"Neocognitron: A self-organizing neural network model for a mechanism of pattern recognition unaffected by shift in position,","cited_arxiv_id":null,"evidence_quote":"Supplies the historical precedent of pre-defined filters in neural networks, motivating the PFCNN idea."},{"cited_title":"CNN Filter DB: An Empirical Investigation of Trained Convolutional Filters,","cited_arxiv_id":null,"evidence_quote":"Provides empirical evidence that trained CNN filters often resemble edge filters, justifying the choice of the 16 edge kernels."},{"cited_title":"ImageNet Classification with Deep Convolutional Neural Networks,","cited_arxiv_id":null,"evidence_quote":"Offers the classic observation that first-layer filters include edge-like detectors, supporting the fixed edge-filter bias."}],"review_version":1}