Adam, Vincent; (2018) Probabilistic models of contextual effects in Auditory Pitch Perception. Doctoral thesis (Ph.D), UCL (University College London).

Preview

Text Adam_10052097_thesis.pdf Download (11MB) | Preview

Abstract

Perception was recognised by Helmholtz as an inferential process whereby learned expectations about the environment combine with sensory experience to give rise to percepts. Expectations are flexible, built from past experiences over multiple time-scales. What is the nature of perceptual expectations? How are they learned? How do they affect perception? These are the questions I propose to address in this thesis. I focus on two important yet simple perceptual attributes of sounds whose perception is widely regarded as effortless and automatic : pitch and frequency. In a first study, I aim to propose a definition of pitch as the solution of a computational goal. Pitch is a fundamental and salient perceptual attribute of many behaviourally important sounds including speech and music. The effortless nature of its perception has led to the search for a direct physical correlate of pitch and for mechanisms to extract pitch from peripheral neural responses. I propose instead that pitch is the outcome of a probabilistic inference of an underlying periodicity in sounds given a learned statistical prior over naturally pitch-evoking sounds, explaining in a single model a wide range of psychophysical results. In two other psychophysical studies I study how and at what time-scales recent sensory history affects the perception of frequency shifts and pitch shifts. (1) When subjects are presented with ambiguous pitch shifts (using octave ambiguous Shepard tone pairs), I show that sensory history is used to leverage the ambiguity in a way that reflects expectations of spectro-temporal continuity of auditory scenes. (2) In delayed 2 tone frequency discrimination tasks, I explore the contraction bias : when asked to report which of two tones separated by brief silence is higher, subjects behave as though they hear the earlier tone ’contracted’ in frequency towards a combination of recently presented stimulus frequencies, and the mean of the overall distribution of tones used in the experiment. I propose that expectations - the statistical learning of the sampled stimulus distribution - are built online and combined with sensory evidence in a statistically optimal fashion. Models derived in the thesis embody the concept of perception as unconscious inference. The results support the view that even apparently primitive acoustic percepts may derive from subtle statistical inference, suggesting that such inferential processes operate at all levels across our sensory systems.

Download activity - last month

Export as XMLCSVJSON

Download activity - last 12 months

Export as CSVXMLJSON

Downloads by country - last 12 months

Export as CSVXMLJSON

Archive Staff Only

View Item