A Computational Model of Dyslexics' Perceptual Difficulties as Impaired Inference of Sound Statistics

Perception is a complex cognitive process in which noisy signals are extracted from the environment and interpreted. It is generally believed that perceptual resolution is limited by internal noise that constrains people’s ability to differentiate physically similar stimuli. The magnitude of this internal noise is typically estimated using the two-alternative forced choice (2AFC) paradigm, which was introduced to eliminate participants’ perceptual and response biases during experiments (Green & Swets, 1966; Macmillan & Creelman, 2004). In this paradigm, a participant is presented with two temporally separated stimuli that differ along a physical dimension and is instructed to compare them. The common assumption is that the probability of a correct response is determined by the physical difference between the two stimuli, relative to the level of internal noise. Performance is typically characterized by the threshold of discrimination, referred to as the Just Noticeable Difference (JND). Thus, the JND is a measure of the level of internal noise such that the higher the JND, the higher the inferred internal noise. However, if the stimuli are highly predictable, perceptual resolution may not be limited by the magnitude of the internal noise. In other words, the assumption of a one-to-one correspondence between the JND and the internal noise may ignore this potential benefit that derives from previous experience. If the internal representation of a stimulus is noisy and hence unreliable, prior expectations should bias the participant against unlikely stimuli. The larger the uncertainty of the measurements, the larger the contribution of these prior expectations is likely to be. The Bayesian theory of inference defines computationally how expectations regarding the probability distribution of stimuli should be combined with the noisy representations of these stimuli in order to form an optimal posterior percept (Knill & Richards, 1996). One limitation of the Bayesian model is that it relies heavily on the assumption that the prior distribution of stimuli is known to the observer. While this assumption may be plausible in very long experiments comprising a large number of trials (e.g., thousands in Körding & Wolpert, 2004) or in experiments utilizing natural tasks (e.g., in reading; Norris, 2006), it is unclear to what extent a rich Bayesian inference is formed when participants have less experience with a task. Here, we studied participants’ patterns of responses on a 2AFC tone discrimination task in relatively short experiments consisting of tens of trials. We found a substantial context effect, whose extent depended on the statistics of the stimuli used in the task and on participants’ internal noise level. Participants’ pattern of behavior was consistent with an “implicit memory” model in 1