CoChat

Investigating Enhancements to Contrastive Predictive Coding for Human Activity Recognition

Open access
Haresamudram, Harish, Essa, Irfan, Plötz, Thomas · 2023 · 2023 IEEE International Conference on Pervasive Computing and Communications (PerCom)

The DOI matches its Crossref record; 65 of the 65 references it cites were found in a registry or an index; 14 of the 14 sources its claims cite were checked for support.

Abstract

The dichotomy between the challenging nature of obtaining annotations for activities, and the more straightforward nature of data collection from wearables, has resulted in significant interest in the development of techniques that utilize large quantities of unlabeled data for learning representations. Contrastive Predictive Coding (CPC) is one such method, learning effective representations by leveraging properties of time-series data to setup a contrastive future timestep prediction task. In this work, we propose enhancements to CPC, by systematically investigating the encoder architecture, the aggregator network, and the future timestep prediction, resulting in a fully con-volutional architecture. Across sensor positions and activities, our method shows substantial improvements on four of six target datasets, demonstrating its ability to empower a wide range of application scenarios. Further, in the presence of very limited labeled data, our technique significantly outperforms both supervised and self-supervised baselines, positively impacting situations where collecting only a few seconds of labeled data may be possible. This is promising, as CPC does not require specialized data transformations or reconstructions for learning effective representations.

Overview

CoChat Summary

TL;DR

The paper systematically enhances Contrastive Predictive Coding (CPC) for wearable human activity recognition by replacing the encoder with strided convolutions, replacing the GRU aggregator with causal convolutions, and performing future-timestep prediction at each timestep. The enhanced CPC improves over prior CPC and other self-supervised baselines on four of six target datasets, especially for locomotion-style activities and limited labeled data.

Key findings

  • Enhanced CPC + MLP achieved test F1-scores of 59.25 ± 1.31 (HHAR), 40.87 ± 0.50 (Myogym), 78.07 ± 0.27 (Mobiact), 89.35 ± 0.32 (Motionsense), 53.79 ± 0.83 (MHEALTH), and 58.19 ± 1.22 (PAMAP2), compared with CPC at 55.59 ± 1.40, 41.03 ± 0.52, 73.44 ± 0.36, 84.08 ± 0.59, 41.03 ± 0.52, and 55.22 ± 0.92, respectively.
  • The paper reports substantial improvements on four of six target datasets (HHAR, Mobiact, Motionsense, MHEALTH); PAMAP2 is similar and Myogym is worse than CPC.
  • Replacing the GRU aggregator with causal convolutions improved HHAR and leg-based datasets, with MHEALTH improving by around 9% and PAMAP2 by around 6%; the new future-prediction task improved four benchmark datasets.
  • With limited labeled data, enhanced CPC generally outperformed other self-supervised methods; for HHAR it was generally best when at least 25 annotated windows were available, and for Mobiact it significantly improved over the Conv. classifier.

Methods

  • Pre-training used the Capture-24 wrist accelerometer dataset (100 Hz, downsampled to 50 Hz, around 2,500 hours, 10% of windows), split by users 90:10 into train/validation; evaluation used six target datasets: HHAR and Myogym (wrist), Mobiact and Motionsense (waist), MHEALTH and PAMAP2 (leg/ankle).
  • Enhanced CPC uses a four-layer 1D convolutional encoder with (32, 64, 128, 256) filters, kernel size (4, 1, 1, 1), stride (2, 1, 1, 1), reflect padding, ReLU, and dropout p = 0.2; a causal convolutional aggregator tuned over (2, 4, 6) blocks with kernel sizes (2, 3, 4, 5, 6, 7), 256 filters, group normalization, ReLU, and skip connections; and future-timestep prediction at each timestep using InfoNCE with negatives from the entire batch.
  • Experiments used 2-second windows, 50% overlap for target datasets, five-fold cross-validation, random search for hyperparameters, 50 training epochs, Adam optimizer, and batch size 256; baselines included Multi-task self-supervision, Autoencoder, SimCLR, CPC, Conv. classifier, DeepConvLSTM, and GRU classifier.

Limitations

  • The paper states that predicting future timesteps at each sample is computationally expensive, though this is somewhat mitigated by the strided encoder; it also notes that CPC and enhanced CPC perform poorly on fine-grained activities such as Myogym compared with SimCLR and Autoencoder.
  • The paper states that increased striding reduces temporal resolution and negatively impacts downstream recognition, and that enhanced CPC can likely be improved with more pre-training data and deeper encoder/aggregator networks.
  • Evident from the text: pre-training uses only wrist accelerometer data from Capture-24, evaluation is limited to accelerometer-based datasets, and hyperparameter search is limited to 20 pre-training combinations and 50 fine-tuning combinations.

Model-written summary. Check it against the paper before relying on it.

Integrity signals

Registry

Crossref · 10.1109/percom56429.2023.10099197

Cites

65 of 65 references verified0 citation errors · 0 not checkable

Statistics

Nothing to recompute.

Verification

Claims the paper makes

10 Valid · 4 Not valid · 1 Not checked · 70 Own results and interpretations

46 more claims

Create a free account to see all 85 claims and every source verdict.

See all claims — free account
  • It may state two separate claims as one. 25
  • Probably not a substantial claim. 9
  • The quote only partly states this claim. 1

Create a free account to see these claims.

Claims other papers make about this one

Each is checked against this paper’s own text: does it say what the citing paper says it does?

Self-check

Percentages, sums and reported differences re-computed from the paper’s own numbers.

Each contradiction between two passages listed here was confirmed by an independent check.

2 of 2

Citations

Cites

Research works printed but not linked: 1

  1. 10.5555/3495724.3497510 · doi:10.5555/3495724.3497510

Cited by

See all 27

Metadata

Authors
Haresamudram, H., Essa, I., & Plötz, T.
Affiliations
Georgia Institute of Technology (US)
Venue
2023 IEEE International Conference on Pervasive Computing and Communications (PerCom)
Publisher
IEEE
Published
2023-03-13
Pages
232-241
ISSN
2474-249X (ISSN-L)
OpenAlex
W4366249647
Cited by
27
References
63
Citations by year
2026: 4 · 2025: 8 · 2024: 13 · 2023: 2
FWCI
2.88
Type
conference paper
Language
en
Open access
Locations
  • IEEE International Conference on Pervasive Computing and Communications (publishedVersion)
  • arXiv (Cornell University) (submittedVersion)
  • arXiv (Cornell University)
Topics
  • Physical Sciences › Computer Science › Computer Vision and Pattern Recognition › Context-Aware Activity Recognition Systems
  • Physical Sciences › Computer Science › Signal Processing › Time Series Analysis and Forecasting
  • Physical Sciences › Computer Science › Artificial Intelligence › Domain Adaptation and Few-Shot Learning
Keywords
contrastive predictive coding, human activity recognition, wearable sensors, self-supervised learning, unlabeled data, encoder architecture, limited labeled data

About these checks

Registry, retraction, Cites and statistics results are computed, not judged, with one exception: a printed reference that no index lists is read once by a model before it counts as a citation error. Claims are extracted by a model and kept only when their quote appears verbatim in the text; each cited source is then checked, with a model’s help, for whether it says what the claim relies on. Those verdicts can be wrong — read the claim’s quote and the cited source before relying on one.

The claim verdicts this report counts are experimental: their accuracy is not yet measured.

A language model drew the claims from the paper on 2026-10-05, under version 3 of the claim definition (model deepseek-v4.1-flash, prompt extract-v3).

A second model labelled each claim on 2026-10-05 (model jev-1.13-20260917, definition revision gold-guideline-v0.2).

Cited sources are checked with a language model’s help (gpt-6-sol).

Statistics and arithmetic are recomputed, not judged.

Confidence bands are shown only for fields calibrated on our annotated test set. None is calibrated yet, so no band is shown.

What counts as a claim ›How we measure confidence ›How we check this ›

Are you an author, or do you see a problem with this page? Request a review

Data sources

Bibliographic data: Semantic Scholar, OpenAlex, Crossref.

Contains information from the Semantic Scholar Academic Graph Datasets, made available under the ODC Attribution License (ODC-BY 1.0).

Semantic Scholar’s data is described in: Kinney, R., et al. (2023). The Semantic Scholar Open Data Platform. arXiv:2301.10140.