Modeling goal-directed attention in tone sequences using a weighted Kalman filter

Humans exhibit a great ability to attend to particular sound sources while ignoring competing streams. Attention is an important process that facilitates focusing computational resources on sound elements of interest in a scene. Attention is believed to facilitate segregation of sound streams by locking onto the characteristics of a `target´ sound; hence giving it more weight compared to other sounds and tracking its evolution over time. In this paper, we explore the hypothesis that the segregation process occurs through tracking target tokens and ignoring the background as outliers to the attended stream. We implement this hypothesis using a weighted Kalman filter approach. The scheme is tested on sinusoidal patterns using classic streaming two tone paradigms. The attentive model developed here is able to attend to a target stream of interest, hence emulating how humans attend to a particular sound in presence of multiple sounds.