Most wearable activity recognition systems assume a predefined sensor deployment that remains unchanged during runtime. However, this assumption does not reflect real-life conditions. During the normal use of such systems, users may place the sensors in a position different from the predefined sensor placement. Also, sensors may move from their original location to a different one, due to a loose attachment. Activity recognition systems trained on activity patterns characteristic of a given sensor deployment may likely fail due to sensor displacements. In this work, we innovatively explore the effects of sensor displacement induced by both the intentional misplacement of sensors and self-placement by the user. The effects of sensor displacement are analyzed for standard activity recognition techniques, as well as for an alternate robust sensor fusion method proposed in a previous work. While classical recognition models show little tolerance to sensor displacement, the proposed method is proven to have notable capabilities to assimilate the changes introduced in the sensor position due to self-placement and provides considerable improvements for large misplacements.

No abstract

In machine learning, concept drift can cause the optimal solution to a given problem to change as time passes, leading to less accurate predictions. Concept drift can be sudden, gradual or reoccuring. Understanding the consequences of concept drift is particularly important in human-centric applications where changes in the underlying data and environment are common and unexpected. In order to gain a better understanding of the adverse effects of different types of concept drift on learners, we propose a novel simulation tool that is able to incrementally generate datasets with customisable concept drift by interacting with a human in a game-like setting. We illustrate our approach by generating and analysing concept drift simulations inspired by body-sensor based long-term activity recognition. Our initial results show that current unsupervised adaptation techniques can be caught in cyclic mislabelling and that a hybrid solution that is selfcalibrating and semi-supervised is more robust than any of the two taken separately for this example.

No abstract

Miller and Licklider's explorations of the intelligibility of temporally interrupted speech, and later studies extending their findings to the spectro-temporal plane, have shown how the twin factors of sparseness and redundancy confer a high degree of robustness on speech in noise. The current contribution addresses two questions. First, to what extent can quantitative estimates of supra-threshold unmasked speech account for average (macroscopic) intelligiblity across a range of speech styles and masking conditions? We examine how well glimpse-based objective intelligibility metrics predict listeners' speech recognition scores for natural and synthetic speech in the presence of stationary and fluctuating maskers, and demonstrate reduced correlations for competing sources with an informational masking component. The second question concerns which additional components, beyond speech glimpses, are required to make (microscopic) predictions of actual listener confusions at the level of individual noisy speech tokens. Using corpora of speech-in-noise misperceptions, we show that in many cases the source of listener confusions is the misallocation of information from the masker, suggesting that estimates of supra-threshold unmasked speech alone are insufficient to explain speech intelligibility in noise.