Abstract
Activity recognition in wearable computing faces two key challenges: i) activity characteristics may be context-dependent and change under different contexts or situations; ii) unknown contexts and situations can lead to incorrect and overconfident predictions. This paper introduces a deep "mixture of experts" architecture (the $\alpha$-$\beta$ network) combined with Maximum Entropy Learning (MEL) to improve human activity recognition from wearable sensors. By learning to identify latent contexts unsupervised and quantifying its own uncertainty, the model not only boosts activity recognition accuracy by up to 10% but can also reliably flag when a user has entered a completely new, unknown context.
Introduction
In wearable computing, recognizing user activities (e.g., walking, typing, eating) is heavily dependent on the environmental context. Traditional context-aware systems often rely on rigidly predefined environments and require extensive labeled contextual data. However, real-world contexts are fluid, and users frequently encounter novel environments. Standard deep learning models are notoriously brittle in these situations, often making highly confident but incorrect predictions. This research addresses these challenges by developing a model that can discover contexts unsupervised, utilize them to improve recognition accuracy, and explicitly detect unknown contexts through uncertainty quantification.
Methods
The proposed framework consists of three primary components designed to handle context-dependent activity recognition and uncertainty:
- The $\alpha$-$\beta$ Network: This Mixture of Experts (MoE) architecture utilizes an $\alpha$ network as a "context detector" to output a probability distribution over latent contexts. Dedicated $\beta$ networks act as "experts" for each context, specializing in activity recognition within that specific environment. The system is trained jointly using an Expectation-Maximization (EM) algorithm.
- Uncertainty Quantification (UQ) via Maximum Entropy Learning (MEL): To detect unknown contexts, MEL is applied to the $\alpha$ network. Instead of fixed weights, MEL learns a distribution of parameters that maximizes entropy, allowing the model to calculate a formal "rejection probability." If uncertainty exceeds a learned threshold, the system flags the context as unknown.
- Clustering-based Pre-training: To prevent "model collapse"—where the gating network routes all data to a single expert—the authors use a CNN to extract features, cluster them via K-means, and pre-train the $\alpha$ network on these clusters.
Key Results
The framework was evaluated on the UCI OPPORTUNITY benchmark and a new, realistic in-house dataset.
- Accuracy Improvements: On the UCI dataset, the $\alpha$-$\beta$ network achieved 96% accuracy using 9 latent contexts, representing a 10% improvement over a standard single-network baseline (86%).
- Unknown Context Detection: The UQ method significantly outperformed probabilistic baselines in detecting held-out, unknown contexts. The proposed method achieved a sensitivity of 0.63–0.75, compared to 0.17–0.34 for logistic regression baselines.
- Real-World Robustness: On the noisier in-house dataset, the model maintained superior performance (84% vs. 80% accuracy) and successfully isolated unknown contexts in its probability distributions.
Discussion and Future Work
The results demonstrate that unsupervised context discovery is a powerful performance booster, simplifying the search space for activity classifiers. Furthermore, the use of parameter distributions via MEL provides a calibrated sense of uncertainty, allowing the model to reject unfamiliar data rather than making incorrect guesses.
While the model showed strong performance, the drop in accuracy on unconstrained real-world data (from 96% to 84%) highlights the ongoing challenge of sensor noise and placement variability. Future work will explore learning multiple levels of contextual information dynamically and investigating additional sources of uncertainty, such as inherent sensor noise.
Summary updated