Number-based human mind state inference in human-machine collaborative systems

In developing the human-machine technology, it is essentially important to infer human mind state. A machine learning approach is promising to this need. However, the machine-learning approach essentially requires training data, ideally supervised training data, which may not be readily available. An idea is to overcome this shortcoming is to take the so-called subjective rate measure. Take the problem of inferring the cognitive fatigue state as an example. This means that we need to ask human subjects to rate their fatigue state while they are performing a task under a particular environment. This is notoriously known problematic as it is intrusive to task performing. In this paper, we propose a notion called “number-based information” as opposed to “word-based information” in terms of applications. We then apply this notion to the problem of mind state inference, leading to a novel inference approach by a combination of physiological signals and task performance. We illustrate this method by using the example of cognitive fatigue inference in the context of rehabilitation for post-stroke patients. Another contribution of this paper is the study of individual-based and group-based strategies to acquire training data to infer the human mind state. In particular, we show a significant improvement in the accuracy of inference with the individual-based strategy.