Temporal Class Activation Mapping for One Dimensional Signals

This direction is merged into another direction: temporal class activation mapping. Its last version stays readable and citable here.

Version 1 of 120 papers22 conceptsfrozen 2026-10-04

Deep learning models have become remarkably adept at classifying complex time-series data, from detecting heart arrhythmias in ECGs to identifying appliance signatures in power grids. However, these models often operate as "black boxes," providing a prediction without explaining which specific part of a long, undulating signal triggered the decision.

To bridge this gap, researchers have adapted Class Activation Mapping (CAM)—a technique originally designed to highlight objects in 2D images—to 1D temporal signals. By projecting the internal state of a neural network back onto the raw input, these methods generate heatmaps that reveal exactly which timestamps or waveforms drive a classification.

From Pixels to Timestamps

In computer vision, CAM identifies which pixels contribute to a label like "cat" or "dog." In the 1D domain, the goal is to identify which temporal segments contribute to labels like "Atrial Fibrillation" or "Sleep Stage N1."

The intuition behind this adaptation relies on the architecture of Convolutional Neural Networks (CNNs). As a signal passes through convolutional layers, the network extracts increasingly abstract features. By using a Global Average Pooling (GAP) layer before the final classification, the model collapses these spatial/temporal features into a single vector.

As Karl Øyvind Mikalsen and Arthur Revhaug demonstrate in Uncertainty-Aware Deep Ensembles for Reliable and Explainable Predictions of Clinical Time Series, we can take the weights of the final layer and multiply them by the feature maps of the last convolutional layer. This produces a "saliency map" that correlates directly with the original time axis, highlighting the most influential moments in the sequence.

Validating Neural Intuition

For clinicians and engineers, these heatmaps serve as a sanity check. If a model predicts a heart attack, a doctor needs to know if the AI is looking at a clinically relevant ST-segment elevation or just electronic noise.

Research by Sebastian D. Goodfellow and colleagues in Towards Understanding ECG Rhythm Classification Using Convolutional Neural Networks and Attention Mappings showed that CAM could successfully highlight the absence of P-waves in ECG signals, a known physiological marker for atrial fibrillation. Similarly, Shivam Kumar Sharma used 1D-GradCAM in Data-efficient Deep Learning Approach for Single-Channel EEG-Based Sleep Stage Classification with Model Interpretability to prove that their model was focusing on K-complexes and alpha waves—the same markers used by human sleep experts.

This alignment between "AI intuition" and domain expertise is critical for adoption in high-stakes environments. For instance, Guangkun Nie applied saliency mapping to photoplethysmography (PPG) signals in Artificial intelligence-derived photoplethysmography age as a digital biomarker for cardiovascular health, ensuring the model captured biological landmarks rather than artifacts when estimating vascular age.

Turning Explanations into Detectors

Perhaps the most powerful evolution of this research is the shift from explaining a classification to localizing events using only "weak" labels.

Traditionally, if you wanted to train a model to find exactly when a toaster turns on in a house's total energy signal, you would need thousands of manually labeled timestamps. However, Adrien Petralia and Themis Palpanas introduced a framework called CamAL in Few Labels are All you Need: A Weakly Supervised Framework for Appliance Localization in Smart-Meter Series.

Instead of precise timestamps, they trained a model using only "possession" labels (e.g., "this house has a toaster"). By extracting the Class Activation Maps from this simple classifier, they could pinpoint the toaster's activity with high accuracy. The "explanation" of the classification effectively became a functional tool for event detection, requiring up to 5,200x fewer labels than traditional supervised methods.

Beyond Simple Classification

The utility of 1D activation mapping extends into anomaly detection and security. In How Do Neural Spoofing Countermeasures Detect Partially Spoofed Audio?, Tianchi Liu and Haizhou Li used Grad-CAM to reveal that spoofing detectors don't just listen for "fake" voices; they specifically hunt for the "seams" or transitions where synthetic audio is spliced into real speech.

Similarly, Dongha Lee and Hwanjo Yu demonstrated in Weakly Supervised Temporal Anomaly Segmentation with Dynamic Time Warping that CAM can generate "pseudo-labels" for anomalous segments in industrial sensor data. This allows models to learn the boundaries of a mechanical failure even when only provided with a general "this sequence contains an error" tag.

By treating interpretability not as a side effect but as a core mechanism, this research direction transforms 1D deep learning from a opaque statistical tool into a transparent, localized, and data-efficient diagnostic framework.

Papers in this version

History & citationVersion 1 · 2026-10-04

Merged into another direction · Permanent link

View frozen version 1

Cite this direction

Latest version (a living page):

@misc{lacuna_temporal_class_activation_mapping_for_on_169e58ab,
  title = {Temporal Class Activation Mapping for One Dimensional Signals},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/temporal-class-activation-mapping-for-one-dimensional-signals/rdr_169e58ab0c1243d9a5b3095555ffb67f},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-20},
}

This exact version (version 1, frozen text):

@misc{lacuna_temporal_class_activation_mapping_for_on_169e58ab_v1,
  title = {Temporal Class Activation Mapping for One Dimensional Signals},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/temporal-class-activation-mapping-for-one-dimensional-signals/rdr_169e58ab0c1243d9a5b3095555ffb67f/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-20},
}