Title: NEST: Neural Estimation by Sequential Testing

Abstract: Adaptive psychophysical procedures aim to increase the efficiency and reliability of measurements. With increasing stimulus and experiment complexity in the last decade, estimating multi-dimensional psychometric functions has become a challenging task for adaptive procedures. If the experimenter has limited information about the underlying psychometric function, it is not possible to use parametric techniques developed for the multi-dimensional stimulus space. Although there are non-parametric approaches that use Gaussian process methods and specific hand-crafted acquisition functions, their performance is sensitive to proper selection of the kernel function, which is not always straightforward. In this work, we use a neural network as the psychometric function estimator and introduce a novel acquisition function for stimulus selection. We thoroughly benchmark our technique both using simulations and by conducting psychovisual experiments under realistic conditions. We show that our method outperforms the state of the art without the need to select a kernel function and significantly reduces the experiment duration.