{"id":"f896ad18-9e60-44b4-9e24-1eb2500808ee","arxiv_id":"1908.05040","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"Deep Generalized Max Pooling, a differentiable global pooling layer based on generalized max pooling, improves CNN embeddings for manuscript classification and writer identification over average and max pooling.","lead":"This paper turns a classic bag-of-words aggregation method, generalized max pooling, into a trainable layer for convolutional neural networks. The new layer re-weights each spatial position's feature vector so frequent and rare patterns contribute more equally, and it beats standard average and max pooling on manuscript and writer identification tasks.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"D>N in both reported settings makes DGMP's linear solve underdetermined, so the layer may be doing exact interpolation instead of GMP's burstiness equalization; the paper offers no diagnostic to rule out unstable weights.","rationale":"The reader's CONDITIONAL verdict remains appropriate. The core mathematics leading to Eq. (9) is standard, and the reported accuracies do support some empirical advantage for DGMP on the three datasets. The reader's weakest assumption flags near-singularity of the Gram matrix and unstable negative weights; my concern is sharper and more specific: in both reported configurations D>N, so the least-squares problem in Eq. (7) is underdetermined. The layer can fit all N spatial constraints nearly exactly, which changes the character of GMP from an approximate equalization over many descriptors to a per-image interpolation. This explains why lambda initialization matters so much and why the manuscript's dimensional remarks are internally inconsistent. The paper does not report condition numbers, residual norms, or alpha statistics for either dataset, so the reader cannot tell whether the descriptor is a stable aggregate or a fragile exact fit. The missing variance estimates for the writer-ID experiments make this concern more acute: a single margin-0.1 run could easily be within seed noise. The concrete test I propose would settle the question by instrumenting the public code and comparing the ridge solution against the unregularized minimum-norm solution. No change to the reader's verdict is needed now; the paper should be accepted only after this diagnostic and a corrected, caveated superiority claim are provided.","tokens_in":11149,"tokens_out":15951,"duration_ms":180855,"concrete_test":"Run the public DGMP implementation on the validation splits of ICDAR17-WI and CLAMM17 and instrument the layer: for each image, record the condition number kappa(K + lambda I), the residual ||K alpha - 1_N||, and the distribution of alpha (min, max, fraction of negative entries, and ||alpha|| relative to 1/N). Then, as a control, replace the ridge solution with the lambda -> 0 minimum-norm solution (Phi^dagger applied to 1_N) on the same features and re-run the writer-ID mAP and CLAMM accuracy. If kappa is high or ||alpha|| >> 1/N with many negative entries, the layer is performing an unstable near-interpolation. If the lambda -> 0 control matches or exceeds DGMP, the learned lambda is not doing the claimed burstiness equalization; if the control is much worse, the regularization is essential and the interpolation concern is mitigated.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim assumes GMP's burstiness equalization transfers from bag-of-words local descriptors to deep CNN activations (Sec. III.B). But in both experimental configurations, D>N, not N≫D: writer-ID uses 400x400 patches and a standard ResNet50 final map of 13x13=169 positions with D=2048; CLAMM uses the stride-reduced 21x21=441 positions with D=2048. Equation (7) is then an underdetermined least-squares problem: the constraint Phi^T xi = 1_N has more unknowns (D) than equations (N), so an exact (or, for lambda small relative to K, near-exact) interpolation exists for every image. This is the opposite of the original GMP regime (N≫D), where ridge regression genuinely compromises between equalizing many descriptors and keeping xi small. Consequently, alpha = (K + lambda I)^{-1} 1 (Eq. 9) can contain large positive and negative entries, and xi can be dominated by cancellation, overfitting each image's spatial positions rather than balancing frequent versus rare descriptors. The paper's own sensitivity to lambda initialization (Sec. IV.C.1: performance 'depends on the proper choice of lambda') and the inconsistent dimensional claims in Sec. III.B ('DN in both cases) are consistent with this instability. If the reported gains rely on this interpolation regime, the broad claim of superiority over average and max pooling is not robustly explained and may not transfer to high-resolution settings with N≫D.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes Deep Generalized Max Pooling (DGMP), a global pooling layer that replaces standard global average or max pooling in CNNs. The layer interprets each spatial location's depth vector of activations as a local descriptor and re-weights these locations by solving a ridge-regularized least-squares problem derived from Generalized Max Pooling (GMP), originally developed for bag-of-words models. The closed-form solution for the weights is given in Eq. (9), and the resulting weighted sum is ℓ2-normalized. The authors evaluate DGMP on three historical document analysis tasks: writer identification/retrieval on ICDAR 17-WI and script type classification on CLAMM 16 and CLAMM 17, using a ResNet-50 backbone with different pooling layers. They report that DGMP outperforms average and max pooling in most of their experiments, and they also compare with mixed pooling, log-sum-exp pooling, and state-of-the-art methods.","tokens_in":11505,"tokens_out":4226,"duration_ms":41895,"significance":"If the central claim holds, DGMP would be a simple, differentiable drop-in replacement for global pooling layers that improves embeddings for structured, repetitive images, with only one extra learnable parameter. The paper has several strengths: the closed-form solution in Eqs. (7)-(9) is mathematically correct and the layer is fully differentiable, allowing end-to-end training; the code is publicly available; CLAMM experiments report means and standard deviations over five runs; and the evaluation includes a range of pooling baselines. However, the significance is moderated by the fact that GMP itself is an existing technique and the novelty is primarily in phrasing it as a neural network layer, and by the empirical results being modest and not universally in favor of DGMP. The theoretical justification connecting DGMP to burstiness equalization is also weakened by the dimensional mismatch between the original GMP setting and the actual CNN activation volumes used, which is a central concern addressed in the major comments.","major_comments":[{"comment":"The dimensional claim in Section III.B is incorrect and load-bearing: the text states 'D < N for the normal ResNet50 version with 400×400 input image sizes and N≈D for the modified ResNet50 version,' but in both experimental settings D=2048 while N=169 (13×13 for 400×400 input) or N=441 (21×21 for the stride-reduced version). Thus D>N in both configurations. The original GMP formulation in [7] is motivated for N≫D, where ridge regression balances many descriptors; in the underdetermined regime the linear system in Eq. (7) can be satisfied exactly (or nearly so for small λ) for each sample, so the layer may be overfitting the spatial positions rather than equalizing frequent versus rare activations. The authors should correct the dimensional statements, provide diagnostics such as the distribution and norm of the learned weights α, the condition number of K+λI, or the residual ‖Φ^T ξ − 1_N‖, and discuss whether the reported gains are attributable to burstiness equalization or to sample-specific interpolation.","section":"Section III.B"},{"comment":"The abstract and contribution list claim that DGMP is 'superior to both average and max pooling' without qualification, but Table I shows a counterexample: for margin 0.5 and λ=10^5, DGMP achieves mAP 45.24 and top-1 64.47, while average pooling achieves 48.60 and 67.19, respectively. The text in Section IV.C.1 later acknowledges that 'the performance of DGMP relative to the average pooling baseline depends on the proper choice of λ,' but this nuance is absent from the abstract and conclusion. The authors should either qualify the superiority claim (e.g., 'for appropriate λ initialization' or 'for small margins') or restrict it to the configurations where it holds. Additionally, no standard deviations or significance tests are reported for the writer-ID experiments in Table I, so it is unclear whether the observed differences are statistically meaningful.","section":"Abstract and Table I"},{"comment":"The transfer of GMP from hand-crafted local descriptors to deep CNN activations is not sufficiently justified. In the original GMP framework, each φ(x) is a local descriptor in a large set (N≫D), and the goal is to reduce the influence of bursty descriptors in the global similarity. In Section III.B, the authors instead define φ_i as the depth vector at one spatial location; with D=2048 and N=169 or 441, the number of equations is far smaller than the number of unknowns. The statement in Section IV.C.2 that 'more activation vectors are pooled and the solution for the optimization problem becomes more reliable' addresses only N, not the ratio N/D. The authors should provide empirical evidence that the weight vector α behaves as in the intended GMP regime, for example by comparing weight statistics on real activations before and after training, or by testing with a lower-dimensional feature map where N≫D can be achieved.","section":"Section III.A and IV.C.2"}],"minor_comments":[{"comment":"There are several typos and inconsistencies: 'fix-sized' in the abstract, inconsistent use of 'CLAMM' vs 'CLaMM', and 'Equ.' vs 'Eq.' in equations. These should be harmonized.","section":"Throughout"},{"comment":"The caption says 'ICDAR 17-WI indicative samples' but the sub-caption references IDs such as '11-3-IMG MAX 1005484'; it would be clearer to explicitly state that these are example patches from the dataset.","section":"Figure 2"},{"comment":"The experimental setup for writer identification omits the batch size and the number of epochs used for validation selection; providing these details would improve reproducibility.","section":"Section IV.C.1"},{"comment":"In Eq. (8), the notation '1n' should be '1_N' for consistency with Eq. (6); also, the subscripts in 'αgmp,λ' are not defined before use.","section":"Equation (8)"},{"comment":"Table II does not report standard deviations or the number of runs for the writer-ID results, unlike the CLAMM tables; adding this information would strengthen the comparison.","section":"Table II"}],"recommendation":"major_revision","confidential_remarks":"The paper is a reasonable incremental contribution but the central claim of universal superiority over average/max pooling is not supported by the full set of reported results. The dimensional error in Section III.B is concerning because it directly affects the theoretical motivation of the method. I would be willing to consider a revised version that addresses the major comments, especially the D>N issue and the qualification of the superiority claim."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short take: this paper is worth a careful read, but the story it tells about why Deep Generalized Max Pooling works is not supported by the numbers in its own setup. The contribution is real: the authors take Murray et al.'s GMP formulation, make it a differentiable layer with a learnable lambda, and evaluate it end-to-end on three historical document benchmarks. The code is public, the CLAMM experiments have five-run means and standard deviations, and the gains over average/max pooling are consistent in the classification tasks. That is a fair, reproducible piece of work.\n\nThe soft spots are in the interpretation. The paper says GMP's burstiness equalization transfers to CNN activations, but the whole GMP derivation assumes many more descriptors than dimensions (N >> D). In both experimental settings here, D=2048 while N=169 (writer-ID, 400x400 patches through a stride-32 ResNet50) or N=441 (CLAMM with the last block stride reduced). That is D>N, not D>D regime) to test whether the interpolation story accounts for the gains.","headline":"Honest adaptation of GMP into a trainable layer with real empirical gains, but the paper misstates its own dimensions and the burstiness-explanation is unverified in the D>N regime.","tokens_in":12045,"tokens_out":2683,"would_cite":false,"duration_ms":25212,"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 introduces Deep Generalized Max Pooling, a trainable global pooling layer that equalizes frequent and rare activation vectors and shows it outperforms global average and max pooling on medieval manuscript classification and…","keywords":["global pooling","deep generalized max pooling","writer identification","script classification","document image analysis","visual burstiness","ridge regression","convolutional neural networks"],"falsifier":"Run the same ResNet-50 and training protocol on a standard object-category benchmark where backgrounds are not bursty; if DGMP does not beat global average pooling there, the claimed superiority is specific to repetitive document imagery rather than a general pooling improvement.","tokens_in":10953,"feed_emoji":"📜","tokens_out":8434,"duration_ms":76756,"temperature":0.7,"pith_summary":"Deep Generalized Max Pooling is a proposed replacement for the final global pooling layer of a convolutional network. Instead of pooling each activation map by itself, it reads the vector of activations at each spatial position, treats that vector as a local image descriptor, and re-weights the positions by solving a ridge-regularized linear system so that frequent and rare activation patterns contribute equally to the final embedding. The paper argues that this counters 'visual burstiness'—background or repeated script dominating the representation—and that it beats both global average pooling and global max pooling on two historical-document tasks: classifying Latin medieval script types and identifying writers. The layer adds a single trainable parameter to the network and can be trained end-to-end with either a softmax or triplet loss.","feed_headline":"Balanced pooling layer beats average and max pooling on manuscripts","feed_subtitle":"It reweights CNN activations so rare and frequent image regions count equally, improving script and writer recognition.","key_machinery":"The central object is the closed-form ridge-regression solve over the Gram matrix of local activation vectors. Generalized max pooling was originally a bag-of-words aggregation that equalizes each descriptor's contribution to the image-to-image similarity. The paper turns it into a network layer by identifying each spatial position's depth vector as a local descriptor. With activation vectors $\\varphi_i$ and Gram matrix $K$, the weights are $\\alpha=(K+\\lambda I)^{-1}\\mathbf{1}_N$, making frequent activation patterns count no more than rare ones; $\\lambda$ interpolates between sum pooling as $\\lambda\\to\\infty$ and the raw least-squares solution as $\\lambda\\to0$. The layer is differentiable, so $\\lambda$ is trained rather than cross-validated, and the authors increase spatial resolution by reducing the last ResNet block's stride to keep the linear solve stable.","core_discovery":"The central claim is that a global pooling layer derived from generalized max pooling produces better image embeddings than the two standard choices. For an activation volume of height $h$, width $w$, and depth $d$, the layer reads each spatial location as a $d$-dimensional vector, forms their Gram matrix $K$, and computes weights $\\alpha=(K+\\lambda I)^{-1}\\mathbf{1}_N$; the pooled vector is the weighted sum $\\xi=\\Phi\\alpha$, normalized to unit norm. Because the closed-form weights are differentiable, $\\lambda$ can be trained by backpropagation, and the layer can sit wherever global average pooling usually sits. On writer retrieval with a triplet loss the layer raises mAP from roughly 50% for average or max pooling to 54.8%, and on the CLAMM 16 and CLAMM 17 script-classification benchmarks it gives consistently higher accuracy than both baselines, including a single-model improvement over prior systems on CLAMM 16.","pith_inferences":["The same burstiness argument should transfer to other domains with repetitive textures, such as medical imaging, histopathology, or scene text, where background patches dominate the representation; this is a direct extension the paper does not test.","Because the Gram-matrix solve couples all spatial positions, the layer's benefit should depend on spatial resolution: reducing the stride, as the paper does for script classification, increases the number of equations and stabilizes the weights, so other architectures may need a similar adjustment.","If the balancing weights are interpreted as a kind of attention, one could combine DGMP with content-based attention; the paper notes this possibility informally, but does not pursue it."],"forward_implications":["Any CNN that ends in global average pooling can swap in this layer at the cost of one extra parameter; the paper demonstrates this on a standard ResNet-50.","The layer makes image embeddings robust to repetitive visual structure, so performance on script classification and writer identification rises without extra training data or multi-crop testing.","Since the weights depend only on the current sample's Gram matrix, the pooling adapts per image, not per class or per dataset, while $\\lambda$ remains the only learned scalar.","DGMP can be combined with the same training losses used for embeddings—the paper uses hard triplet mining for writer identification and softmax for script classification—suggesting it is a drop-in pooling layer for metric-learning pipelines."],"supporting_citations":[{"why":"supplies the generalized max pooling formulation whose ridge-regression weight solution the paper turns into a network layer.","marker":"[7]"},{"why":"introduced generalized max pooling and is compared against democratic aggregation in the motivation for equalizing descriptor contributions.","marker":"[10]"},{"why":"defines visual burstiness, the phenomenon the layer is designed to counteract.","marker":"[8]"},{"why":"provides the ResNet-50 architecture that hosts the pooling layer in all experiments.","marker":"[4]"},{"why":"introduced global average pooling, the baseline layer DGMP replaces.","marker":"[13]"},{"why":"defines the CLAMM 16 manuscript classification dataset and competition protocol used for evaluation.","marker":"[19]"},{"why":"defines the CLAMM 17 dataset and provides the T-DeepCNN baseline to which the paper compares.","marker":"[20]"},{"why":"defines the Historical-WI'17 writer identification dataset and the triplet-network baseline to which the paper compares.","marker":"[31]"}],"fun_headline_variants":["Deep generalized max pooling beats avg and max on manuscripts","Reweighted CNN pooling lifts manuscript and writer recognition","Pooling that balances rare and frequent boosts script ID","Generalized max pooling outperforms standard pooling layers","New pooling layer reweights activations for better embeddings"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the vector of CNN activations at one spatial location behaves like a local descriptor in a bag-of-words model, so equalizing frequent and rare such vectors improves deep image embeddings; if deep activation vectors are too highly correlated for the Gram-matrix solve to produce stable weights, the pooling advantage could vanish.","fun_headline_variants_meta":{"raw":{"variants":["Deep generalized max pooling beats avg and max on manuscripts","Reweighted CNN pooling lifts manuscript and writer recognition","Pooling that balances rare and frequent boosts script ID","Generalized max pooling outperforms standard pooling layers","New pooling layer reweights activations for better embeddings"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000445,"raw_usage":{"total_tokens":2221,"prompt_tokens":884,"completion_tokens":1337,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":500,"completion_tokens_details":{"reasoning_tokens":1272}},"tokens_in":500,"tokens_out":1337,"duration_ms":9546,"temperature":1.0,"reasoning_tokens":1272,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T13:25:15.497650+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the same ResNet-50 and training protocol on a standard object-category benchmark where backgrounds are not bursty; if DGMP does not beat global average pooling there, the claimed superiority is specific to repetitive document imagery rather than a general pooling improvement.","supporting_citations":[{"cited_title":"Interferences in Match Kernels,","cited_arxiv_id":null,"evidence_quote":"supplies the generalized max pooling formulation whose ridge-regression weight solution the paper turns into a network layer."},{"cited_title":"Generalized Max Pooling,","cited_arxiv_id":null,"evidence_quote":"introduced generalized max pooling and is compared against democratic aggregation in the motivation for equalizing descriptor contributions."},{"cited_title":"On the Burstiness of Visual Elements,","cited_arxiv_id":null,"evidence_quote":"defines visual burstiness, the phenomenon the layer is designed to counteract."},{"cited_title":"Deep Residual Learning for Image Recognition,","cited_arxiv_id":null,"evidence_quote":"provides the ResNet-50 architecture that hosts the pooling layer in all experiments."},{"cited_title":"ICFHR2016 Competition on the Classification of Medieval Handwritings in Latin Script,","cited_arxiv_id":null,"evidence_quote":"defines the CLAMM 16 manuscript classification dataset and competition protocol used for evaluation."},{"cited_title":"ICDAR2017 Competition on the Classification of Medieval Handwritings in Latin Script,","cited_arxiv_id":null,"evidence_quote":"defines the CLAMM 17 dataset and provides the T-DeepCNN baseline to which the paper compares."},{"cited_title":"ICDAR2017 Competition on Historical Document Writer Identification (Historical-WI),","cited_arxiv_id":null,"evidence_quote":"defines the Historical-WI'17 writer identification dataset and the triplet-network baseline to which the paper compares."}],"review_version":1}