Gaussian process classification for psychophysical detection tasks in multiple populations (wide big data) using transfer learning

Abstract

Psychophysics is the diagnostic mechanism used to measure neuronal responses to physical or non-physical stimuli, and therefore a means to understand human cognition and behavior. In recent years, Machine Learning (ML) methods have been introduced to speed up and make more efficient very costly psychophysical detection tasks (PDTs) as part of cognitive neurosciences and other similar fields. However, most ML-driven PDT tasks depend on data from single populations despite the growing sets of PDT databases from other populations. This paper therefore offers a brief tutorial on the non-parametric Gaussian process classification (GPC) as a Bayesian machine learning inference method that is growing in appeal for ML-driven PDTs and can be used across multiple populations (wide Big Data) using transfer learning. We then perform a systematic review of GPC in audiogram estimation as an example. GPC reduced expensive diagnostic PDTs in audiometry from as many as a minimum of 400 trials to less than 16 trials, in some instances extending to both ears at the same time. GPC has also significantly been influenced by active learning to maximize PDT experimental data and trials, making them faster and more accessible to wider audiences outside the lab, for example with digital devices such as mobile or web-based apps. GPC with active learning also improved PDTs to measure finer stimulus intensities compared with wider traditional intensity intervals. We conclude by offering suggestions on how GPC combined with active and transfer learning presents an opportunity to serve as an ideal Big Data approach to ML-driven PDTs across multiple populations. We offer some suggestions and areas for further research in other diagnostic assessments.