This chapter describes an adaptive machine learning (ML) method for the utilization of unlabeled data for continual model adaptation after deployment. Current methods for the usage of unlabeled data, such as unsupervised and semi-supervised methods, rely on being both smooth and static in their distributions. In this chapter, a generic method for leveraging causal relationships to automatically associate labels with unlabeled data using state transitions of asynchronous interacting cause and effect events is discussed. This self-labeling method is predicated on a defined causal relationship and associated temporal spacing. The theoretical foundation of the self-supervised method is discussed and compared with its contemporary semi-supervised counterparts using dynamical systems theory. Implementations of this method to adapt action recognition ML models in semiconductor manufacturing and human assembly tasks as manufacturing cyber-physical systems (CPS) are provided to demonstrate the effectiveness of the proposed methodology.
Top1. Introduction
Adaptive machine learning (ML) methods aim to equip deployed machine learning models with the domain adaptability to counter domain and data distribution shifts (Kouw & Marco, 2019; Lu et al., 2018). This allows models to continuously update and improve over time in dynamic, ever-changing environments. There are several types of data distribution shifts, including covariate shift (Schneider et al., 2020), label shift (Garg et al., 2020), and concept drift (Lu et al., 2018). Among these three types of data distribution shifts, concept drift is of great interest to be studied as it affects the entire input-output relation, compared to other drifts that affect either input or output distributions. Concept drift severely degrades the performance of trained models as system nonstationarity invalidates learned relationships (Yang et al., 2019). Therefore, post-deployment adaptability is imperative for ML models to sustain in dynamic environments. In general, ML models can be updated via retraining with datasets that include recent data distribution changes to learn and adapt to said changes (Lu et al., 2018). Generating updated datasets typically requires manual data collection and annotation, which, while providing precise and high-quality data can be costly and laborious (Fredriksson et al., 2020).
To ease the problem of data annotation cost, a class of methods for adaptation to data distribution shifts known as semi-supervised learning (SSL) was developed. SSL applies learned relationships from an initial dataset to newly obtained target data (Van Engelen & Hoos, 2020). These semi-supervised methods can be applied to unsupervised domain adaptation (UDA) using labeled data from the initial distribution (Wilson & Cook, 2020). Semi-supervised methods include pseudo-labels, where unlabeled data is used to generate target class labels used as pseudo-labels to guide adaptation (Lee, 2013), and label propagation, a graph-based approach to apply labels to unlabeled data (Zhu & Ghahramani., 2002). Similar semi-supervised label generation methods mainly use generative networks or clustering methods and generally assume target distribution overlap (Yang et al., 2023). Efforts have been made to adapt to greater domain shifts (Zhang et al., 2020), but they still rely on the correlation between initial and target domains. UDA, as it does not incorporate target labels, is bounded by the discrepancy and joint error of the distributions (Zhang et al., 2023). Recently, several works have proposed to jointly optimize pseudo-label generation and target learning (Asano et al., 2019; Zhou et al., 2021; Yan et al., 2021). Overall, the pseudo-label-based methods show promising results in reducing manual data annotation, but the post-deployment data distribution shifts can degrade their performance in dynamic environments. Causality-based semi-supervised learning has recently attracted researchers’ attention due to the intrinsic invariance of causal relationships across domains (Schölkopf, 2022). Most research in this direction relies on the assumption that the generation of cause data and the causal mechanism are statistically independent (Gong et al., 2016). This chapter was inspired by the causality-based adaptive methods and introduces an adaptive ML system that considers the temporal relationships in causality.