On the Impact of Representation Dimensionality in Self-Supervised Learning for Human Activity Recognition

The official repository of the paper accepted at SIBGRAPI 2026. The study evaluates three SSL models (HARSCnnEncoder and two CNNPFF variants) trained on three SSL techniques (DIET, LFR, and TNC) across seven latent space dimensionalities (9, 45, 90, 180, 360, 720, and 768) and six datasets from the DAGHAR benchmark (Ku-HAR, MotionSense, RealWorld-Thigh, RealWorld-Waist, UCI-HAR, and WISDM).

This project was developed as part of the Cognitive Architectures research line from the Hub for Artificial Intelligence and Cognitive Architectures (H.IAAC) of the State University of Campinas (UNICAMP). See more projects from the group here.

Repository Structure

All our analysis are made per dataset. This repository provides three csv files with the exploration results, four analysis jupiter notebook files, and several figures.

Summary csv files:

• summary-accuracy.csv: A summary of the accuracies achieved by our evaluated models.
• summary-clustering.csv: A summary of the clustering quality of the representations learned by the evaluated pretrained models.
• summary-backbone-macs-params.csv: The number of trainable parameters and total MACs of the backbones explored. Values were computed using the thop library.

Analysis jupyter notebook files:

• analysis-cluster.ipynb: Comparison and ranking of the pretrained backbones, based on their cluster quality.
• analysis-pareto.ipynb: Comparison of the evaluated backbones in terms of model size and MACs. Additionally, points in the Pareto frontier were identified using the accuracies in summary-accuracy.csv.
• analysis-statistical-tests.ipynb: Statistical tests comparing different sets of observations. It leverages the accuracies in summary-accuracy.csv, the Wilcoxon signed-rank test, and the graphviz library for renderization.
• analysis-top-results.ipynb: Maximum accuracies across three backbones, three SSL techniques, and seven latent space dimensionalities.

Figures folder:

• figures/ : All figures employed in the study. In figures/latent-space-visualization, we present a list of files visualizing the learned representations by the best and worst model configurations for DIET, LFR, and TNC.

Dependencies / Requirements

The following libraries are required for the execution of the analysis:

• Minerva
• ultralytics-thop
• graphviz

Authors

• (2026) Darlinne Hubert Palo Soto: PhD student, IC-Unicamp

Acknowledgements

This work was supported in part by the Ministry of Science, Technology, and Innovation of Brazil, with resources granted by Federal Law 8.248 of October 23, 1991, through the PPI Program administered by Softex (PPI-Softex) under Grant 01245.003479/2024-10; in part by CAPES; in part by CNPq under Grants 315399/2023-6 and 302458/2022-0; and in part by FAPESP under Grants 2013/08293-7 and 2023/12865-8.