Information Preserving Downsampling for Temporal and Spiking Signals

18 concepts10 papers1 proposal

In convolutional neural networks designed for computer vision, max-pooling is a foundational operation. It provides translation invariance and extracts the most prominent features—such as a sharp edge or a distinct corner—by selecting the maximum value within a local neighborhood.

However, when architectures shift from static images to temporal sequences, continuous waveforms, or discrete spike trains, this peak-centric selection becomes an information-destroying bottleneck. Max-pooling discards the precise timing (phase), the aggregate density of events, and low-amplitude variations. To combat this, a growing body of research focuses on information-preserving downsampling techniques that maintain the structural and temporal integrity of complex signals.

The Destructive Nature of Max-Pooling on Waveforms

For periodic and continuous signals—such as speech, electrocardiograms (ECG), or electroencephalograms (EEG)—the exact shape and timing of a wave carry critical diagnostic and semantic information. When a standard max-pooling window slides across these signals, it acts as an aggressive non-linear filter that erases phase relationships.

Max-Pooling (Destructive):
Raw Wave:    /\_/\_/\_/\   -->  Extract Peaks Only  -->  [Max, Max, Max]  (Phase & Shape Lost)

Average Pooling (Preserving):
Raw Wave:    /\_/\_/\_/\   -->  Integrate Area     -->  [Mean, Mean, Mean] (Context & Amplitude Retained)

As noted by Baek et al., standard downsampling techniques discard the critical phase and timing information inherent in periodic biological signals, which is essential for tasks like cuffless blood pressure estimation. When max-pooling eliminates these temporal relationships, deeper layers cannot reconstruct the underlying dynamics of the signal.

In speech recognition, this loss manifests as a structural bottleneck. Coppieters de Gibson et al. demonstrated that max-pooling operations destroy the phase information required for higher-level auditory processing. This destruction forces early layers to prematurely aggregate features and learn suboptimal, wide-band filter distributions to compensate, rather than naturally synthesizing complex hierarchical representations in deeper layers.

To preserve these subtle variations, researchers often substitute max-pooling with average or mean pooling:

  • EEG Signal Reconstruction: Wei et al. utilized 1D average pooling instead of max pooling to remove muscle artifacts from EEG recordings. Because average pooling smooths the signal representation across the temporal field, it preserves low-amplitude neural data and subtle waveform details that peak-focused methods would discard, allowing the network to separate overlapping neural and myogenic components.
  • Medical Perfusion Mapping: In computed tomography perfusion imaging, Gava et al. substituted average pooling for max pooling to preserve subtle intensity variations across temporal sequences, ensuring that the continuous flow of contrast agents is accurately mapped.

Smoothing the Discrete: Mean Pooling in Spiking Networks

The information loss caused by max-pooling is even more pronounced in Spiking Neural Networks (SNNs). SNNs communicate via discrete, binary spike trains over time. If a max-pooling operation is applied to a binary window containing spikes, the presence of a single spike outputs a 1.

This creates a severe mathematical blind spot: a window with one spike yields the exact same output as a window packed with ten spikes. The network completely loses its ability to represent spike density or rate-based information.

Binary Spike Train Window A:  [0, 0, 1, 0]  --> Max Pooling = 1
Binary Spike Train Window B:  [1, 1, 1, 1]  --> Max Pooling = 1  (Information Loss)

Binary Spike Train Window A:  [0, 0, 1, 0]  --> Mean Pooling = 0.25
Binary Spike Train Window B:  [1, 1, 1, 1]  --> Mean Pooling = 1.00 (Density Preserved)

To mitigate this, mean pooling is widely preferred in convolutional spiking architectures. By averaging binary events across a neighborhood, mean pooling converts discrete spikes into continuous, graded representations.

As Cheng et al. explain, mean pooling prevents the significant information loss associated with max pooling in discrete spike trains, preserving a more representative signal of the underlying temporal dynamics. This preservation is vital for maintaining structural context. For instance, in energy-efficient 3D point cloud denoising, Li et al. showed that mean pooling prevents the loss of geometric detail inherent in binary signals by converting discrete events into continuous averages that retain the underlying structural context. Similarly, Kucik et al. found that substituting max-pooling with average pooling ensures that the aggregate activity of a local receptive field is propagated, maintaining the representational integrity of the spiking signal.

Architectural Strategies: Delaying and Decomposing

Beyond changing the pooling function itself, researchers have designed structural strategies to safeguard fragile temporal data, focusing on when and how downsampling occurs.

Delayed Downsampling

One straightforward method to protect fine-grained details is to postpone dimensionality reduction. In large vocabulary continuous speech recognition, Bi et al. successfully scaled convolutional networks by delaying pooling and stride-based dimensionality reduction until deeper layers. Keeping the resolution high during initial feature extraction prevents internal data representations from shrinking too rapidly and preserves narrow-band spectral details.

Pooling vs. Striding

When downsampling is necessary, the choice of mechanism matters. In deep learning-based side-channel analysis (such as analyzing power or electromagnetic traces of cryptographic hardware), Karayalcin et al. demonstrated that pooling layers should be preferred over strided convolutions. Strided convolutions skip across data points with a fixed step size, which risks bypassing localized, high-frequency leakage points entirely. Pooling, by contrast, aggregates information across the entire stride window, ensuring these critical leakage points are not omitted.

Learnable Frequency Decomposition

A more advanced alternative to traditional pooling is to treat downsampling as a signal decomposition problem. Inspired by the "Lifting Scheme" in signal processing, Hu et al. introduced Temporal Lift Pooling (TLP) for continuous sign language recognition.

Temporal Lift Pooling:
                      /--> Predict High-Frequency Details (Movements)
Raw Temporal Signal --
                      \--> Update Low-Frequency Approximations (Structure)

Instead of using a static mathematical function like a maximum or an average, TLP decomposes temporal signals into high-frequency details and low-frequency approximations using a three-stage learnable process (splitting, predicting, and updating) via 1D convolutions. This allows the network to explicitly separate fine, rapid movements from core structural information, preserving highly discriminative temporal features that standard, non-learnable pooling operations would otherwise discard.

Key Foundational Work

Very deep convolutional neural networks for LVCSR (2015) — contributed 2 concept elements to this direction.
End-to-End Blood Pressure Prediction via Fully Convolutional Networks (2019) — contributed 1 concept elements to this direction.
History & citationVersion 2 · 2026-10-04

Current · Permanent link

  1. Version 2 (latest) · 2026-10-04 · 10 papers
  2. Version 1 · 2026-10-04 · 11 papers
Cite this direction

Latest version (a living page):

@misc{lacuna_information_preserving_spike_pooling_c3dedce9,
  title = {information-preserving spike pooling},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/information-preserving-spike-pooling/rdr_c3dedce9a5c54189b52bdf7415c0fff6},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-20},
}

This exact version (version 2, frozen text):

@misc{lacuna_information_preserving_spike_pooling_c3dedce9_v2,
  title = {information-preserving spike pooling},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/information-preserving-spike-pooling/rdr_c3dedce9a5c54189b52bdf7415c0fff6/v/2},
  note = {Version 2, frozen 2026-10-04. Accessed 2026-11-20},
}
About

These directions overlap with this one through shared papers.

See something wrong? Report a mistake