{"id":"e9499bc9-9140-4c6c-a458-5ea0e9555cf8","arxiv_id":"2412.16751","paper_version":2,"verdict":"REJECT","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"high","formal_verification":"none","parameter_count":0,"one_line_summary":"Depthwise separable CNN spatial filters transfer well even from the last layers across datasets and architectures, but the paper's evidence is weakened by missing baseline controls.","lead":"This paper tests whether the visual filters inside a certain kind of convolutional network stay general in deep layers, and reports that they do. It shows that freezing and reusing filters trained on one image dataset works even in the last layers, challenging an older result that deep filters become specialized.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The paper never compares frozen trained filters against frozen random or fixed-basis filters, so the transfer results do not distinguish learned master keys from arbitrary depthwise kernels.","rationale":"I agree with the reader's weakest assumption: the missing frozen-random-filter control is the most load-bearing gap, because the central claim is about trained filters being special rather than arbitrary kernels being sufficient. The paper's own shuffle and first-three-layer-repeated results already suggest that exact trained filter identities can be permuted or replaced without loss, which makes the 'master key' interpretation fragile. The selffer inconsistency (Pets 36.3% original vs 52.4% selffer) further weakens the quantitative claims, but the random-filter arm is the decisive experiment for the hypothesis. I therefore recommend keeping the reader's REJECT verdict; if the random arm underperforms clearly, the paper's central claim would gain real support.","tokens_in":10550,"tokens_out":8455,"duration_ms":73982,"concrete_test":"Run ConvNeXt Femto with the paper's 300-epoch recipe on Oxford Pets and on the ImageNet man-made/natural split under three frozen-depthwise conditions: (a) trained source filters as in the paper, (b) Xavier/Gaussian random kernels with identical shape and per-layer norms, and (c) fixed hand-designed basis filters (e.g., Gabor/Laplacian) repeated across channels. Use at least 5 seeds per condition and report mean with standard error. If (b) or (c) matches (a) within roughly 1-2 top-1 points, the trained filter identity is not the driver and the Master Key Filters Hypothesis is unsupported; if both clearly underperform, the concern is resolved.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The Master Key Filters Hypothesis claims trained depthwise filters converge to universal 'master key' sets, and that this convergence is why frozen filters transfer. The experiments in 'Generality of Spatial Features in DS-CNNs' and Tables 1-4 never test the null hypothesis that arbitrary frozen depthwise kernels, such as random, Gabor, or constant filters, would transfer equally well once the pointwise layers are trained. The controls that are present instead weaken the hypothesis: shuffling filters across layers and repeating only the first three layers' filters (Table 1) preserve accuracy, showing that exact trained filter identity and layer placement are not what matters. Since ConvNeXt's depthwise convolution sits between LayerNorm and pointwise layers, pointwise layers could plausibly compensate for any spatial kernel. A second, related confound is the selffer baseline: Pets selffer accuracy is 52.4% versus the original 36.3% (Tables 3 and 4), so the baseline used to measure 'improvement' is not matched and conflates the self-transfer protocol with longer or easier optimization of the pointwise layers. Without a frozen-random-filter arm, the data are equally consistent with 'trained filter identity is irrelevant' as with 'trained filters converge to master keys.'","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"This paper proposes the Master Key Filters Hypothesis: that depthwise-separable convolutional filters converge to universal \"master key\" filter sets and remain generic across layers, datasets, and architectures. The authors support this with transfer experiments in which depthwise filters are frozen while the pointwise layers are trained: a semantically split ImageNet replication on ConvNeXt and ResNets, a six-dataset cross-domain transfer matrix on ConvNeXt Femto, cross-architecture transfers from ConvNeXt and HorNet to ConvNeXt Femto, and a combined cross-domain/cross-architecture transfer. They report that, unlike in Yosinski et al. (2014), transferring deeper layers does not degrade accuracy, and that pointwise-filter transfer degrades performance.","tokens_in":10614,"tokens_out":8518,"duration_ms":73673,"significance":"If established, the hypothesis would be a notable challenge to the classical layer-specialization picture and would have practical implications for transfer learning and architecture design. The paper collects a large transfer matrix and includes several architectures, which is a useful empirical starting point. However, the experiments do not currently discriminate between \"trained DS-CNN filters converge to universal master keys\" and the much weaker statement \"arbitrary frozen spatial kernels work because the pointwise layers can compensate\"; the self-transfer baseline also conflates additional pointwise optimization with the value of the source filters. The manuscript therefore does not yet establish the hypothesized mechanism, although the underlying question is worth pursuing.","major_comments":[{"comment":"The central inference that frozen trained filters are intrinsically general is missing the crucial control arm of frozen random, fixed-basis (e.g., Gabor), or constant spatial filters. Since the pointwise layers are always trained from scratch after freezing, high transfer accuracy could be achieved by any spatial kernel that the pointwise layers can compensate for. This concern is sharpened by the paper's own Table 1: shuffling filters across layers (86.2%) and repeating only the first three layers' filters in every layer (86.9%) essentially match the baseline (86.9%), showing that exact trained filter identity and depthwise-layer assignment are not necessary for the reported transfer performance. A frozen-random-filter arm must be added before the results can support convergence to master keys.","section":"Generality of Spatial Features in DS-CNNs, Tables 1 and 3-4"},{"comment":"The \"selffer\" baseline is not matched with the original training protocol. For Oxford Pets the selffer run reaches 52.4% while the original 300-epoch run reaches 36.3% (Tables 3 and 4); similarly, STL10 and Flowers improve from 80.4% to 82.7% and from 66.0% to 69.1%. Because selffer re-trains the pointwise layers for another 300 epochs on the same target-domain data, it combines frozen-filter transfer with additional optimization of the pointwise/classifier parameters. Improvements attributed to source filters should instead be measured against a control with the same number of pointwise update steps but no source filter transfer, such as random frozen filters or a model trained from scratch for the same total budget.","section":"Cross Domain Transfer, Experimental Setup and Tables 3-4"},{"comment":"The hypothesis is not operationalized. The paper never defines what a \"master key filter set\" is, how to enumerate one, or what observable prediction would falsify it. The experiments use the same DS-CNN families whose filter-clustering observations motivated the hypothesis, and the hypothesis is asserted after the fact. To make it testable, the authors should specify a concrete candidate master-key set, for example cluster centers from one architecture trained on one dataset, and test a priori whether those exact filters transfer better than arbitrary filters; alternatively, they should state a predicted quantitative relationship between filter distances and transfer accuracy.","section":"Introduction (Master Key Filters Hypothesis)"},{"comment":"The generality claim is broader than the evidence. The cross-domain transfer matrix uses a single small architecture (ConvNeXt Femto), the architecture comparison covers only ConvNeXt and HorNet, and all tasks are natural-image classification benchmarks. The abstract's statement that spatial features \"remain generic across all layers, domains, and architectures\" is therefore not supported. Either the claim should be restricted to the tested settings or the experiments should be extended to additional architecture families and task types.","section":"Abstract and Conclusions and Discussion"},{"comment":"The numerical presentation of the central transfer tables is ambiguous. Table 4's caption describes diagonal cells as selffer accuracies while off-diagonal entries appear to be relative changes, and the arrow definition refers to comparisons \"compared to the selffer models in each row\" even though columns are the target datasets. Without a clear statement of units (absolute accuracy versus percentage-point change) and baselines, the reader cannot verify the claims of asymmetry and dataset-size effects. The authors should report absolute accuracies and the corresponding baselines in a single consistent layout, or provide both.","section":"Tables 4 and 5"}],"minor_comments":[{"comment":"Table 1 reports single runs without error bars or multiple seeds; differences of 0.7 percentage points are presented as support for equality, so the reader cannot assess whether the differences are significant.","section":"Table 1"},{"comment":"Please copyedit for typographical issues, including \"frozendepthwise\" in the Table 4 caption, \"scalabilty\" in Related Work, and \"it's\" in the Figure 3b text.","section":"Throughout"},{"comment":"Figure 6's caption and the surrounding text are difficult to reconcile with Table 4: the text says transfer from Foods improves performance on Pets, but Table 4 lists Food-to-Pets as a negative value if read as a change, and the figure legend does not state the baseline to which the curves are compared.","section":"Figure 6"}],"recommendation":"major_revision","confidential_remarks":"I think the empirical transfer matrix is worth developing, but the central hypothesis cannot be evaluated without random-filter controls and a properly matched selffer baseline. If the authors are unwilling to add those experiments, I would recommend rejection; as it stands, the manuscript is a useful descriptive study rather than a test of the Master Key Filters Hypothesis. No code or training seeds are mentioned, which makes reproducibility difficult to assess."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The genuinely new thing here is the empirical claim: frozen depthwise filters from deep layers transfer across dissimilar datasets and architectures without the degradation Yosinski et al. reported for traditional CNNs. That is a real datapoint, and the cross-architecture transfer from HorNet to ConvNeXt is a nice addition. The paper is clearly written and the experimental sweep is broad within its chosen scope. If all they had claimed was 'DS-CNN depthwise filters transfer better than Yosinski's AlexNet-style filters,' I'd be mostly convinced.\n\nThe soft spots are real and load-bearing. Most importantly, there is no random-filter control. The paper tests shuffling filters across layers and repeating the first three layers' filters, and accuracy holds. That already tells you the exact identity and layer placement of trained filters are not what matters. Without comparing against frozen random or fixed kernels (Gabor, constant), the results are equally consistent with 'any fixed depthwise kernel works because pointwise layers compensate.' The Master Key Filters Hypothesis needs that control to distinguish learned generality from architectural redundancy.\n\nThe second problem is the selffer baseline. Pets selffer accuracy is 52.4% versus the original 36.3% in Table 3. That gap suggests the self-transfer protocol is not a matched baseline—it includes retraining pointwise layers for another 300 epochs, which could simply be extra optimization. The claimed improvements in Tables 4 and 7 are measured against that weak baseline, so the magnitude of the effect is overstated.\n\nI also agree with the reader that the abstract's 'across all layers, domains, and architectures' is broader than the evidence: only two architecture families, one small target model, and natural-image classification datasets. And the hypothesis itself has no operational definition or falsifiable prediction—'master key sets' are never enumerated in advance.\n\nThe stress-test note holds up on reading the paper. The missing random arm is not a manufactured concern; the paper's own shuffle and repeat experiments point in that direction.\n\nThat said, the paper is not incoherent. It engages honestly with Yosinski and with its own prior clustering work. The empirical puzzle—why deep depthwise filters transfer when standard ones don't—is worth pursuing. A serious referee should see it, but with a demand for a random-filter arm, a properly matched selffer baseline, and more restrained claims.\n\nRecommendation: send to peer review. The core observation deserves scrutiny, and the missing controls are fixable.","headline":"Fresh empirical counterpoint to Yosinski for depthwise nets, but the missing random-filter arm and an unmatched selffer baseline leave the 'master key' claim unproven.","tokens_in":11286,"tokens_out":1984,"would_cite":false,"duration_ms":19870,"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":"This paper argues that depthwise separable convolution filters converge toward universal 'master key' filter sets, and shows via frozen-filter transfer experiments that these spatial filters stay general across all layers, domains, and…","keywords":["depthwise separable convolution","master key filters","filter transferability","layer specialization","cross-domain transfer","cross-architecture transfer","generalization in neural networks","ConvNeXt"],"falsifier":"Run the same cross-domain transfer protocol with randomly initialized depthwise filters frozen in place; if the random-filter model matches the trained-filter accuracy on Oxford Pets within a point or two, the central claim is not specifically about learned filters, whereas a large gap would confirm that learned filters carry the generality.","tokens_in":10220,"feed_emoji":"🔑","tokens_out":9714,"duration_ms":78194,"temperature":0.7,"pith_summary":"The paper proposes the Master Key Filters Hypothesis: deep networks built from depthwise separable convolutions learn spatial filters that are universal, with reusable patterns appearing in every layer, on every dataset, and in every architecture tested. To support this, it freezes depthwise filters from a source model and trains the remaining network on a target dataset, across a semantically split ImageNet, six unrelated datasets, and different model families. The transfers hold up: accuracy does not degrade as more layers are frozen, and filters from larger datasets can even improve target performance. If true, this challenges the long-held picture that later layers become class-specific and points to depthwise filters as a reusable visual vocabulary.","feed_headline":"Deep filters are universal keys across visual tasks","feed_subtitle":"Frozen spatial filters from one dataset or architecture work in another, challenging the deeper-is-specialized view.","key_machinery":"The load-bearing object is the depthwise separable convolution, which splits a standard convolution into a per-channel spatial filter (depthwise) followed by a 1x1 channel mixer (pointwise). This split makes the spatial component directly visible and transferable: a depthwise kernel is a small 2D pattern applied to one channel, and the paper freezes exactly these kernels while retraining everything else. The experimental machinery is the frozen-filter transfer protocol from the classic layer-transfer study, applied layer-by-layer and in a 'selffer' variant where the model transfers filters to itself, plus a semantically divided ImageNet (man-made versus natural) as a demanding dissimilar-domain test. Together these let the paper isolate whether spatial filters are the general part of the network.","core_discovery":"The paper's central claim is that the spatial features encoded by depthwise separable convolutions are generic rather than specialized: the depthwise filters of DS-CNNs converge to 'master key filter sets' that work for visual data generally, independent of the training dataset, task, or architecture. The evidence is a sequence of frozen-filter transfer experiments: in the man-made/natural ImageNet split, ConvNeXt retains full accuracy even when the last depthwise layer is transferred; across six datasets, filters from larger source datasets help regardless of domain similarity; and filters transfer across architectures such as HorNet to ConvNeXt, including when both domain and architecture differ. In contrast, freezing transferred pointwise filters consistently hurts, which the paper attributes to optimization difficulties such as fragile co-adaptation rather than to specialization of those filters. The paper therefore concludes that the depthwise component is the general spatial operator while the pointwise component carries adaptation to the task.","pith_inferences":["The paper's claim would be stronger with a control that freezes randomly initialized depthwise filters; without it, the reported success could be explained by the trainable pointwise layers compensating for whatever spatial kernels are present, rather than by the learned filters themselves.","The consistent dataset-size gradient suggests a testable corollary: pretraining spatial filters on larger, more diverse unlabeled data should improve downstream transfer more than matching the target domain, making master-key filters a scaling phenomenon rather than a domain-specific one.","If the filters are truly universal operators resembling Gaussian derivatives, the same master keys might transfer to non-image modalities such as audio spectrograms or time series, an extension the paper does not explore."],"forward_implications":["Transferring depthwise filters from a trained DS-CNN is a viable initialization for a new visual task, with the deepest layers transferring without the drop seen in traditional CNNs.","The flat transfer curve across depth contradicts the assumption that later layers are class-specific; DS-CNNs do not exhibit the specialization gradient reported for traditional CNNs.","Pointwise layers carry task adaptation, so design efforts to specialize a model should target the pointwise mixing rather than the spatial kernels.","Filter sets trained on large generic data act as reusable visual priors, so small datasets can benefit from source models trained on unrelated large datasets.","Cross-architecture transfer is viable even when block designs differ, as HorNet-to-ConvNeXt results show, enabling modular reuse of spatial filter stacks."],"supporting_citations":[{"why":"Supplies the layerwise frozen-filter transfer methodology and the opposing claim that deeper filters specialize, which this paper directly re-runs on DS-CNNs.","marker":"Yosinski et al. 2014"},{"why":"Reports clusterable repeating patterns in ImageNet-trained depthwise kernels, the empirical motivation for the Master Key Filters Hypothesis.","marker":"Babaiee et al. 2024a"},{"why":"Introduces depthwise separable convolutions, the architecture whose depthwise filters this paper isolates and transfers.","marker":"Howard et al. 2017, 2019"},{"why":"Defines the ConvNeXt family used as the base model and as source and target in most transfer experiments.","marker":"Liu et al. 2022"},{"why":"Provides HorNet, a structurally different DS-CNN family used to test cross-architecture transfer.","marker":"Rao et al. 2022"}],"fun_headline_variants":["Deep filters are master keys, not specialists","Frozen deep filters transfer across domains and architectures","DS-CNN deep filters stay generic, transfer like master keys","Why deep filters are actually universal keys","Deep filters: generic across tasks, not class-specific"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing assumption is that a model's accuracy with frozen transferred filters measures how general those filters are; the paper never freezes random or trivial filters, so it cannot rule out that the trainable layers can compensate for almost any spatial kernel.","fun_headline_variants_meta":{"raw":{"variants":["Deep filters are master keys, not specialists","Frozen deep filters transfer across domains and architectures","DS-CNN deep filters stay generic, transfer like master keys","Why deep filters are actually universal keys","Deep filters: generic across tasks, not class-specific"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000187,"raw_usage":{"total_tokens":1289,"prompt_tokens":868,"completion_tokens":421,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":484,"completion_tokens_details":{"reasoning_tokens":349}},"tokens_in":484,"tokens_out":421,"duration_ms":4023,"temperature":1.0,"reasoning_tokens":349,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T10:14:44.821367+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the same cross-domain transfer protocol with randomly initialized depthwise filters frozen in place; if the random-filter model matches the trained-filter accuracy on Oxford Pets within a point or two, the central claim is not specifically about learned filters, whereas a large gap would confirm that learned filters carry the generality.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the layerwise frozen-filter transfer methodology and the opposing claim that deeper filters specialize, which this paper directly re-runs on DS-CNNs."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides HorNet, a structurally different DS-CNN family used to test cross-architecture transfer."}],"review_version":1}