Assessing the State of Self-Supervised Human Activity Recognition Using Wearables

Abstract

The emergence of self-supervised learning in the field of wearables-based human activity recognition (HAR) has opened up opportunities to tackle the most pressing challenges in the field, namely to exploit unlabeled data to derive reliable recognition systems for scenarios where only small amounts of labeled training samples can be collected. As such, self-supervision, i.e., the paradigm of 'pretrain-then-finetune' has the potential to become a strong alternative to the predominant end-to-end training approaches, let alone hand-crafted features for the classic activity recognition chain. Recently a number of contributions have been made that introduced self-supervised learning into the field of HAR, including, Multi-task self-supervision, Masked Reconstruction, CPC, and SimCLR, to name but a few. With the initial success of these methods, the time has come for a systematic inventory and analysis of the potential self-supervised learning has for the field. This paper provides exactly that. We assess the progress of self-supervised HAR research by introducing a framework that performs a multi-faceted exploration of model performance. We organize the framework into three dimensions, each containing three constituent criteria, such that each dimension captures specific aspects of performance, including the robustness to differing source and target conditions, the influence of dataset characteristics, and the feature space characteristics. We utilize this framework to assess seven state-of-the-art self-supervised methods for HAR, leading to the formulation of insights into the properties of these techniques and to establish their value towards learning representations for diverse scenarios.

Contrastive Predictive Coding for Human Activity Recognition

“…Such design requires domain expertise and results in representations that are useful for downstream applications. This modeling paradigm of 'pretrain-thenfinetune' has resulted in significant performance boosts for human activity recognition from wearables [19], [27], [30].…”

“…Another contrastive learning method includes SimCLR [24], [26], [27], which heavily relies on data transformations [32] as it contrasts two randomly augmented views of the data in a siamese setup, resulting in highly effective representations. SimSiam [27], [33] and Bootstrap Your Own Latent (BYOL) [27], [34] are also siamese methods, albeit do not contrast the two views. For multi-sensor setups, ColloSSL [35] and COCOA [36] contrast across different sensor locations for training.…”

“…In contrast, Myogym contains fine-grained gym activities such as different types of curls, extensions, raises, rows, etc., with the null class covering over 75% of the data. As in [27], evaluating across different sensor positions and activities allows us to perform a broad evaluation of the proposed method and understand its strengths and areas of opportunity.…”

“…The pretext task involves future timestep prediction in a contrastive learning setup, which leverages properties of time-series data (e.g., the ability to predict future timesteps from the past) for effective training. Such a framework is attractive as it does not require the design of specialized data transformations (like in the case of SimCLR [26] for example), which can be dependent on the data streams utilized, or reconstruction of (parts of) the input data (e.g., Autoencoders [23] and Masked reconstruction [20]), which may not perform well under distribution shifts [27].…”

“…Specifically, we study the encoder architecture, the autoregressive network, and the future prediction tasks of the original setup in order to derive suitable alternatives that can improve HAR performance. As in [27], we perform pre-training on the largescale Capture-24 dataset, which contains accelerometer data collected at the wrist. The evaluation is performed on six diverse datasets, where two datasets each contain sensor data from the wrist, waist, and the leg, so as to enable the study of transferring weights across sensor locations.…”