Learned Downsampling with Strided Convolutions

21 concepts18 papers1 proposal

In classical convolutional neural network (CNN) architectures, spatial downsampling has long been dominated by pooling operations. Max-pooling and average-pooling layers reduce the spatial dimensions of feature maps by applying fixed, non-parametric heuristics across a local neighborhood. While computationally cheap, pooling operations discard a massive amount of information—a standard $2\times2$ max-pooling layer blindly throws away 75% of its activations, retaining only the maximum value.

This aggressive reduction introduces a fundamental bottleneck: fixed pooling layers cannot adapt to the specific information preservation needs of a given task. To overcome this limitation, a prominent research direction replaces rigid pooling layers with strided convolutions, transforming downsampling from a hard-coded heuristic into a learnable optimization problem.

The Intuition: Why Learnable Downsampling Works

To understand why strided convolutions outperform traditional pooling, consider how they process an image. A max-pooling layer acts as a deterministic filter that prioritizes high-amplitude activations regardless of context, which frequently destroys fine-grained spatial relationships, phase information, and subtle gradients.

A convolution with a stride greater than one, however, possesses learnable weights. During backpropagation, the network optimizes these weights to function as task-specific downsampling filters. For example, the network can learn to implement a smooth low-pass filter to prevent aliasing, or an edge-detecting filter that selectively preserves boundaries. Instead of blindly dropping pixels, a strided convolution mathematically blends and compresses spatial information, allowing the model to retain the exact features necessary for downstream tasks.

Streamlining Architectures and Optimizing Invariance

The foundational argument for eliminating pooling layers altogether was established by Alexey Dosovitskiy et al. in Striving for Simplicity: The All Convolutional Net. The authors demonstrated that a "pure" network consisting solely of convolutional layers could match or exceed the performance of traditional architectures on image recognition tasks. They showed that the spatial invariance required for object recognition does not need to be forced via rigid pooling functions; instead, the network can dynamically learn the optimal downsampling transformations.

Building on this philosophy, modern architectures have increasingly embraced non-pooling designs to streamline feature extraction. For instance, Weihu Song et al. introduced a medical image segmentation framework in Non-pooling Network for medical image segmentation that completely eliminates pooling operations. By relying entirely on strided convolutions, their network captures hierarchical features with significantly reduced information loss, proving that learnable downsampling can simultaneously lower computational parameters while boosting accuracy.

Preserving Spatial Precision for Localization

For tasks where the exact geometric arrangement of pixels is critical—such as image segmentation, patch matching, and edge detection—the spatial blurring caused by max-pooling can be highly destructive. Strided convolutions preserve spatial alignment by maintaining a continuous, learnable mapping across layers.

This precision is highly beneficial across several domains:

  • Patch Matching: In L2-Net: Deep Learning of Discriminative Patch Descriptor in Euclidean Space, Yurun Tian et al. utilized strided convolutions instead of pooling to maintain structural information within intermediate feature maps. This architectural choice ensures that the precise spatial correspondences required for high-performance descriptor matching are not lost during dimensionality reduction.
  • Line and Feature Detection: When localizing fine-grained linear structures, traditional pooling often erases thin boundaries. To solve this, Van Nhan Nguyen et al. designed LS-Net: fast single-shot line-segment detector to replace max-pooling layers with strided convolutions. The authors demonstrated that preserving fine-grained spatial information through learnable downsampling is critical for the precise regression of linear features.
  • Medical Ultrasound Segmentation: In ultrasound-guided surgery, maintaining exact positional information is a matter of safety. In Simultaneous Segmentation and Classification of Bone Surfaces from Ultrasound Using a Multi-feature Guided CNN, Puyang Wang et al. substituted max-pooling with strided convolutions in a U-Net architecture, successfully preserving the fine spatial details and positional context necessary to accurately map bone surfaces.

Mitigating the Vanishing Target Problem

A major defect of fixed pooling is its tendency to erase small-scale features. If a critical target or pattern occupies only a tiny fraction of pixels in a local neighborhood, a max-pooling operation can easily overwrite it with a more dominant, irrelevant neighbor activation.

This issue is particularly acute in medical imaging and remote sensing, where targets are often minute. To combat the "vanishing small target" problem, Liangli Xiong et al. developed SEA-NET: medical image segmentation network based on spiral squeeze-and-excitation and attention modules. By replacing non-parametric pooling with strided convolutions, the network learns how to compress feature maps without discarding the subtle pixel information indicative of small targets, such as overlapping cells in blood smears.

Similarly, in the domain of satellite imagery, Bai Liu et al. applied this strategy in LAND COVER CLASSIFICATION USING CONVOLUTIONAL NEURAL NETWORK WITH REMOTE SENSING DATA AND DIGITAL SURFACE MODEL. They found that avoiding standard pooling layers maintains higher spatial resolution in the deeper layers of the network, which is essential for defining sharp object boundaries and precise land cover localization.

Application to 1D Signals and Network Efficiency

The intuition behind learned downsampling naturally extends beyond 2D computer vision into 1D signal processing and compact network design.

In medical signal analysis, Hsin-Tien Chiang et al. deployed strided convolutions within a fully convolutional denoising autoencoder in Noise Reduction in ECG Signals Using Fully Convolutional Denoising Autoencoders. Because heart signals rely on precise temporal alignment and subtle waveform shapes, strided convolutions allow the network to compress the signal for storage while preserving the fine-grained structural details required for accurate clinical denoising. For audio security, Zhenzong Wu et al. demonstrated in Light Convolutional Neural Network with Feature Genuinization for Detection of Synthetic Speech Attacks that learnable downsampling preserves the intricate temporal and spectral patterns necessary to differentiate genuine human speech from synthetic deepfake audio.

Crucially, replacing pooling with strided convolutions does not require sacrificing network efficiency. Nazanin Beheshti explored this trade-off in Squeeze U-Net: A Memory and Energy Efficient Image Segmentation Network. By substituting fixed pooling layers with learnable strided convolutions, the architecture optimizes its parameters to retain maximum spatial information during training, achieving a massive reduction in model size and energy consumption without compromising segmentation accuracy.

Key Foundational Work

Striving for Simplicity: The All Convolutional Net (2015) — contributed 2 concept elements to this direction.
L2-Net: Deep Learning of Discriminative Patch Descriptor in Euclidean Space (2017) — contributed 1 concept elements to this direction.
History & citationVersion 1 · 2026-10-04

Current · Permanent link

View frozen version 1

Cite this direction

Latest version (a living page):

@misc{lacuna_strided_convolution_downsampling_8a2c477d,
  title = {strided convolution downsampling},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/strided-convolution-downsampling/rdr_8a2c477d82dd4a9999be9557b0ee5219},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-22},
}

This exact version (version 1, frozen text):

@misc{lacuna_strided_convolution_downsampling_8a2c477d_v1,
  title = {strided convolution downsampling},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/strided-convolution-downsampling/rdr_8a2c477d82dd4a9999be9557b0ee5219/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-22},
}
About

These directions overlap with this one through shared papers.

See something wrong? Report a mistake