Hybrid Pooling Downsampling (HPD)

• Hybrid Pooling Downsampling (HPD) is a family of techniques that integrate diverse pooling methods to preserve critical features during resolution reduction.
• HPD is applied in CNNs for image classification and segmentation, and in graphs for topology-preserving coarsening with multi-scale structures.
• HPD methods combine operators such as max, average, min pooling, and spectral techniques to optimize information retention with modest computational overhead.

Searching arXiv for the cited HPD-related papers to ground the article and citations. Hybrid Pooling Downsampling (HPD) denotes a family of downsampling strategies that replace fixed, single-criterion reduction with hybrid operators intended to preserve more task-relevant information during resolution reduction. The term is not used canonically across the literature: it has been used for a hybrid of DiffStride and Spectral Pooling in convolutional networks (Rafif et al., 2024), for graph-level readout operators that fuse attention, sum, and max pooling in mixup-based graph learning (Dong et al., 2022), for a MinMaxPooling-based encoder module in medical image segmentation (Yue et al., 20 Jul 2025), and more broadly for a design viewpoint in which downsampling mixes pooling criteria, multi-scale structure, or pre-filtering (Gholamalinezhad et al., 2020). This suggests that HPD is best understood as an architectural principle—hybridization of downsampling mechanisms to improve information retention—rather than as a single fixed operator.

1. Terminological scope and conceptual basis

Pooling layers are described as the component that executes down-sampling on feature maps, drastically reduces spatial dimension, reduces the number of parameters or weights, lessens computational cost, and controls overfitting (Gholamalinezhad et al., 2020). Within that general role, HPD reframes downsampling as a learnable or structured aggregation problem rather than a fixed application of max pooling, average pooling, or constant stride.

Setting	HPD meaning	Representative source
CNN design viewpoint	mixing multiple pooling mechanisms, multi-scale priors, or pre-filtering	(Gholamalinezhad et al., 2020)
CNN image classification	combined implementation of Spectral Pooling and DiffStride on ResNet-18	(Rafif et al., 2024)
Graph learning	fusion of multiple graph readouts into a graph-level embedding	(Dong et al., 2022)
Semantic segmentation	MinMaxPooling with lightweight feature filtering	(Yue et al., 20 Jul 2025)

In regular CNNs, the motivation is recurrent: fixed stride samples on a rigid grid, max pooling discards submaximal responses, and average pooling can blur edges or suppress salient activations (Rafif et al., 2024, Bieder et al., 2021, Gholamalinezhad et al., 2020). In graph settings, the motivation is analogous but expressed at the level of node-set compression or topology-preserving coarsening, where the objective is to reduce graph size while preserving semantic information or structural properties (Dong et al., 2022, Bacciu et al., 2022). Across these usages, “hybrid” may mean mixing max and average behavior, combining learned stride with spectral truncation, fusing minimum and maximum local responses, or coupling structural coarsening with feature pooling.

2. Regular-domain formulations in convolutional models

A central CNN instantiation appears in "Hybrid of DiffStride and Spectral Pooling in Convolutional Neural Networks" (Rafif et al., 2024). The starting point is standard strided convolution,

$y[n] = \sum_k w[k]\,x[n\cdot s + k],$

where fixed $s$ can skip informative features. In that formulation, DiffStride learns stride parameters $S=(S_h,S_w)$ by backpropagation through a frequency-domain downsampling operator. The output size is coupled to the learned stride as

$$H' = \lfloor H / S_h + 2R \rceil,\qquad W' = \lfloor W / S_w + 2R \rceil,$$

with $R$ a smoothness factor. Spectral Pooling is defined by transforming to the frequency domain, cropping low frequencies to a target size, and inverting:

$$y=T(x),\qquad \hat y = CropSpectrum(y; h\times w),\qquad \hat x=T^{-1}(\hat y).$$

The hybrid places DiffStride in early and mid downsampling locations of ResNet-18 and inserts Spectral Pooling one level above global average pooling. The authors report that placing Spectral Pooling above GAP is more effective than placing a second DiffStride there (Rafif et al., 2024).

A second convolutional formulation appears in "A Novel Downsampling Strategy Based on Information Complementarity for Medical Image Segmentation" (Yue et al., 20 Jul 2025). Here HPD is a MinMaxPooling module that extracts complementary bright and dark information. For input $X\in\mathbb{R}^{C\times H\times W}$, the hard local statistics are

$$y_{\max}(c,i,j)=\max_{(u,v)\in W(i,j)} X(c,u,v),\qquad y_{\min}(c,i,j)=\min_{(u,v)\in W(i,j)} X(c,u,v),$$

followed by either weighted fusion,

$$y(c,i,j)=\alpha\,y_{\max}(c,i,j)+\beta\,y_{\min}(c,i,j),$$

with $\alpha\ge 0$, $s$0, $s$1, or concatenation plus a $s$2 convolution, batch normalization, and ReLU. The paper also gives optional soft min/max approximations through LogSumExp to alleviate gradient sparsity (Yue et al., 20 Jul 2025).

A third regular-domain formulation is provided by AdaPool, explicitly described as an HPD instantiation (Stergiou et al., 2021). AdaPool fuses two exponential kernels per pooling window: an exponentiated Dice–Sørensen-coefficient weighting and an exponential maximum. Its regional fusion is

$s$3

where $s$4 is a learned regional mask. The stated objective is to avoid average pooling’s oversmoothing and max pooling’s extreme selection while retaining non-zero gradients for all positions (Stergiou et al., 2021).

The broader review literature places these methods within a continuum of hybridization strategies. Mixed pooling, rank-based pooling, weighted pooling, adaptive $s$5 pooling, anti-aliased blur pooling, spatial pyramid pooling, and spectral pooling are all treated as mechanisms that tune the information-loss versus invariance trade-off, often by interpolating between average-like and max-like behavior or by adding pre-filtering before decimation (Gholamalinezhad et al., 2020). By contrast, "Comparison of Methods Generalizing Max- and Average-Pooling" restricts HPD to local window operators that interpolate between max and average, including Gated Pooling, Ordinal Pooling, Learned Norm Pooling, and Smooth-Maximum Pooling (Bieder et al., 2021).

3. Graph-based formulations and structural generalization

In language-processing GNNs, HPD refers to hybrid graph pooling operators used as global readout rather than spatial reduction in a grid. "On the Effectiveness of Hybrid Pooling in Mixup-Based Graph Learning for Language Processing" defines node updates and graph readout as

$s$6

The study uses global readout without top-$s$7 selection or soft cluster assignment. Its single-criterion readouts are attention pooling, sum pooling, and max pooling, and the hybrid operators fuse two readouts by elementwise sum, Hadamard product, or concatenation plus linear projection:

$s$8

$s$9

Nine hybrid operators are evaluated. Type 1, $S=(S_h,S_w)$0, and Type 3, $S=(S_h,S_w)$1, are repeatedly identified as strong choices, while Type 6, $S=(S_h,S_w)$2, is singled out in some robustness settings (Dong et al., 2022).

A structurally different graph interpretation appears in "Generalizing Downsampling from Regular Data to Graphs" (Bacciu et al., 2022). There HPD can be instantiated as topology-preserving graph coarsening paired with feature pooling. The coarsening selects centroids through a maximal $S=(S_h,S_w)$3-independent set, induces clusters around them, contracts edges between clusters, and pools node features inside each cluster. The parameter $S=(S_h,S_w)$4 acts as a stride or resolution control: centroids are at least $S=(S_h,S_w)$5 hops apart, every node is within $S=(S_h,S_w)$6 hops of a centroid, and larger $S=(S_h,S_w)$7 yields coarser representations (Bacciu et al., 2022).

This graph-coarsening view is notable because it provides formal correspondence to regular downsampling on diagonal grid graphs. With lexicographic ranking, the paper states that $S=(S_h,S_w)$8-MIS selects the same positions as strided sampling with stride $S=(S_h,S_w)$9, and that the induced partition matches pooling windows of size $$H' = \lfloor H / S_h + 2R \rceil,\qquad W' = \lfloor W / S_w + 2R \rceil,$$0 (Bacciu et al., 2022). In this usage, HPD is hybrid because it combines structural downsampling and feature pooling rather than because it fuses multiple scalar pooling operators.

4. Empirical performance across tasks

The empirical record for HPD is heterogeneous rather than uniformly positive. In the DiffStride–Spectral Pooling hybrid on ResNet-18, the reported mean accuracy on CIFAR-10 increases from 0.924 for the DiffStride baseline to 0.9334 for the hybrid, an improvement of 0.0094, and the CIFAR-100 mean accuracy increases from 0.706 to 0.7382, an improvement of 0.0322 (Rafif et al., 2024). The same study reports that a two-DiffStride combination reaches 0.9282 on CIFAR-10, below the hybrid with Spectral Pooling (Rafif et al., 2024).

In graph learning for language and code, hybrid readouts are reported to outperform MAXPOOL and GMT in both accuracy and robustness. On NLP tasks, Type 1 improves accuracy versus GMT by up to 4.38% on Politifact, and on Gossipcop it achieves average 96.56% versus GMT 95.71%. On code datasets, Type 1 outperforms GMT by up to 2.36% accuracy. Robustness gains are larger in some perturbation settings: Type 6 surpasses GMT by up to 23.23% on GIN with Gossipcop-Profile, and Type 1 improves robustness by up to 10.23% on Python800. The paper further reports that Type 1 is significantly better than No Aug, MAXPOOL, and GMT across most models for accuracy, and in 10/15 tested cases for robustness under Wilcoxon signed-rank testing with Bonferroni correction (Dong et al., 2022).

In semantic segmentation, the MinMaxPooling formulation reports gains on ACDC and Synapse across several backbones, with average improvement of about 0.5% mDSC across architectures (Yue et al., 20 Jul 2025). Illustrative cases include UNet on ACDC, 0.8993 to 0.9032; LinkNet on ACDC, 0.8690 to 0.8767; Fast-SCNN on ACDC, 0.7753 to 0.7863; and Swin-UNet on Synapse, 0.6227 to 0.6531. The paper also reports negative or marginal cases, including Deeplabv3+ on ACDC, 0.8001 to 0.7901, and Fast-SCNN on Synapse, 0.6487 to 0.6465 (Yue et al., 20 Jul 2025).

AdaPool reports systematic gains across multiple vision tasks. On ImageNet classification, the paper gives top-1 improvements such as ResNet-18 from 69.76 to 71.78, ResNet-50 from 76.15 to 78.42, DenseNet-121 from 74.65 to 77.29, ResNeXt-50 32×4d from 77.62 to 79.98, and Wide-ResNet-50 from 78.51 to 80.24. On MS COCO, Mask R-CNN with ResNet-101 is reported to improve AP from 38.2 to 42.4. The same work also reports gains in video action recognition, image super-resolution, and frame interpolation (Stergiou et al., 2021).

By contrast, "Comparison of Methods Generalizing Max- and Average-Pooling" reports that none of the more sophisticated pooling methods perform significantly better than standard max- or average-pooling in VGG16 on Google Open Images v5 (Bieder et al., 2021). Test accuracy values cluster near 39% for AP, NN, OP, LNP, and SEMP, while SMP variants are lower. This paper is an important counterexample to any claim that hybrid pooling is intrinsically superior.

A different form of HPD, defined as dynamic interleaving of full-resolution and low-mode passes with downsampled inputs and pooled kernels during training, reports efficiency rather than accuracy gains. On CIFAR-10, ResNet-18 shows 18% total time reduction with top-1 error moving from 6.7% to 6.9%, and ResNet-50 shows 23% total time reduction with error moving from 6.3% to 6.4% (Poulos et al., 2019). This indicates that HPD can also be used as a training-time efficiency device rather than only as an inference-time architectural module.

5. Theoretical properties, information retention, and computational trade-offs

The main theoretical argument across CNN formulations is information preservation under downsampling. In the DiffStride–Spectral formulation, fixed stride is said to miss features because it samples on a rigid grid, while max pooling imposes hard selection and a constraining lower bound on preserved information; spectral truncation is motivated by retaining low-frequency energy that carries global structure, and learnable stride adapts the sampling rate to the data (Rafif et al., 2024). The same paper references prior work by Rippel et al. for the claim that spectral pooling reduces the constraint on the lower bound of preserved information for a given output dimensionality, although explicit bounds are not proved there (Rafif et al., 2024).

The MinMaxPooling variant bases its theory on “information complementarity”: maximum values emphasize bright structures and strong edges, minimum values emphasize dark structures, baseline, and depressions, and their fusion encodes local contrast and detail (Yue et al., 20 Jul 2025). The paper explicitly notes that hard min/max yields sparse gradients, whereas soft min/max with LogSumExp gives dense gradients at the cost of approximation bias (Yue et al., 20 Jul 2025). AdaPool advances a related but distinct argument: both of its exponential kernels produce strictly positive normalized weights, so all positions receive non-zero gradients, and the learned $$H' = \lfloor H / S_h + 2R \rceil,\qquad W' = \lfloor W / S_w + 2R \rceil,$$1 adapts locally between a smooth-average-like and smooth-max-like regime (Stergiou et al., 2021).

The graph-coarsening formulation provides the clearest structural guarantees. The coarsened graph preserves the number of connected components, and for any coarsened edge $$H' = \lfloor H / S_h + 2R \rceil,\qquad W' = \lfloor W / S_w + 2R \rceil,$$2 with $$H' = \lfloor H / S_h + 2R \rceil,\qquad W' = \lfloor W / S_w + 2R \rceil,$$3 the original graph distance satisfies

$$H' = \lfloor H / S_h + 2R \rceil,\qquad W' = \lfloor W / S_w + 2R \rceil,$$4

For centroid pairs in the same component, the paper further gives path-length bounds

$$H' = \lfloor H / S_h + 2R \rceil,\qquad W' = \lfloor W / S_w + 2R \rceil,$$5

These statements formalize distortion control under graph downsampling (Bacciu et al., 2022).

The computational trade-offs are equally consistent across papers. FFT and Hartley transforms introduce additional cost relative to max or average pooling in the DiffStride–Spectral method (Rafif et al., 2024). MinMaxPooling requires approximately two pooling passes plus optional $$H' = \lfloor H / S_h + 2R \rceil,\qquad W' = \lfloor W / S_w + 2R \rceil,$$6 convolution, and the paper reports modest or small FLOP and parameter increases depending on the backbone (Yue et al., 20 Jul 2025). AdaPool reports minor additional parameter and FLOP overhead but clearly higher pooling latency than average pooling, with CUDA timings of 14 ms / 76 ms for avg pooling forward/backward versus 119 ms / 490 ms for AdaPool (Stergiou et al., 2021). The graph coarsening method emphasizes $$H' = \lfloor H / S_h + 2R \rceil,\qquad W' = \lfloor W / S_w + 2R \rceil,$$7 work using parallel neighborhood reductions without explicitly materializing the dense $$H' = \lfloor H / S_h + 2R \rceil,\qquad W' = \lfloor W / S_w + 2R \rceil,$$8 graph (Bacciu et al., 2022). The training-time HPD scheme predicts reduced per-iteration cost through simultaneous input downsampling and kernel pooling, with empirical low-mode/full-mode cost ratios around 0.37–0.39 for ResNet-18/50 at downsampling factor $$H' = \lfloor H / S_h + 2R \rceil,\qquad W' = \lfloor W / S_w + 2R \rceil,$$9 (Poulos et al., 2019).

6. Architectural integration, limitations, and recurring design choices

Architecturally, HPD is usually inserted at canonical downsampling sites. In ResNet-18, the DiffStride–Spectral variant replaces strided convolutions in shortcut downsampling blocks, applies DiffStride to both main and residual branches for shape alignment, and places Spectral Pooling one level above GAP (Rafif et al., 2024). In UNet-like encoders, the MinMaxPooling variant replaces each downsampling operation with HPD using the standard $R$0, $R$1 pattern, while keeping skip connections unchanged (Yue et al., 20 Jul 2025). In graph classification pipelines, hybrid pooling is applied after $R$2 message-passing layers to produce graph-level embeddings, and Manifold-Mixup is performed at that graph-level representation rather than on adjacency structure (Dong et al., 2022). In the graph-coarsening framework, HPD naturally occupies the coarsen-and-pool step between GNN blocks (Bacciu et al., 2022).

Several recurring design choices appear across the literature. Small or standard reduction factors are preferred: $R$3, $R$4 in segmentation encoders (Yue et al., 20 Jul 2025), $R$5 in dynamic input and kernel downsampling (Poulos et al., 2019), and moderate $R$6 in graph coarsening for a balance between compression and distortion (Bacciu et al., 2022). Hybrid fusion is also typically lightweight: convex combinations, concatenation plus $R$7 projection, or per-region scalar masks rather than large auxiliary networks (Dong et al., 2022, Stergiou et al., 2021, Yue et al., 20 Jul 2025).

The limitations are equally recurrent. Statistical significance is not reported for the DiffStride–Spectral experiments (Rafif et al., 2024). The graph-learning study explicitly notes threats to external and construct validity, including dependence on UPFD and Project_CodeNet and evaluation robustness based on random edge dropping rather than other perturbation models (Dong et al., 2022). The segmentation formulation warns that min pooling can be sensitive to outlier noise and that architectures already rich in multi-scale fusion, such as DeepLabv3+, may experience redundancy (Yue et al., 20 Jul 2025). The comparison study on generalized pooling finds no significant advantage over standard max or average pooling in a large-scale image-classification setting, and observes that freely learned temperatures in Smooth-Maximum Pooling can destabilize training or drift toward overly sharp regimes (Bieder et al., 2021). AdaPool, while broadly successful, still adds measurable runtime overhead (Stergiou et al., 2021).

Taken together, these results rule out a simple interpretation of HPD as a universally dominant replacement for conventional downsampling. The literature instead supports a narrower conclusion: hybridization is beneficial when its chosen components match the information bottleneck of the task—frequency truncation for late CNN representations, min–max complementarity for medical image boundaries, adaptive exponential weighting for detail-preserving vision pipelines, or fused graph readouts for mixup-based graph classification. Where that match is weak, standard pooling may remain competitive or preferable.