Neural network models of categorical perception

Abstract

Studies of the categorical perception (CP) of sensory continua have a long and rich history in psychophysics. In 1977, Macmillan,Kaplan, and Creelman introduced the use of signal detection theory to CP studies. Anderson and colleagues simultaneously proposed the first neural model for CP, yet this line of research has been less well explored. In this paper, we assess the ability of neural-network models of CP to predict the psychophysical performance of real observers with speech sounds and artificial! novel stimuli. We show that a variety of neural mechanisms are capable of generating the characteristics of CPO Hence, CP may not be a special mode of perception but an emergent property of any sufficiently powerful general learning system. Studies of the categorical perception (CP) of sensory continua have a long and rich history. For a comprehensive review up until a decade ago, see the volume edited by Hamad (1987). An important question concerns the precise definition ofCP. According to the seminal contribution of Macmillan, Kaplan, and Creelman (1977), "the relation between an observer's ability to identify stimuli along a continuum and his ability to discriminate between pairs of stimuli drawn from that continuum is a classic problem in psychophysics" (p. 452). The extent to which discrimination is predictable from identification performance has long been considered the acid test of categorical-as opposed to continuous-perception.Continuous perception is often characterized by (approximate) adherence to Weber's law, according to which the difference limen is a constant fraction of stimulus magnitude. Also, discrimination is much better than identification: I Observers can make only a small number of identifications along a single dimension, but they can make relative (e.g., pairwise) discriminations between a much larger number of stimuli (G. A. Miller, 1956). By contrast, CP was originally defined (e.g., Liberman, Cooper, Shankweiler, & Studdert-Kennedy, 1967;Liberman, Harris, Hoffman, & Griffith, 1957) as occurring when the grain of discriminability coincided with (and We are grateful to David Wykes, who programmed the brain-state-ina-box neural simulation. Neil Macmillan kindly clarified for us some technical points relating to the psychophysics of the ABX task. The perceptive, critical comments of Robert Goldstone, Stephen Grossberg, Keith Kluender, Neil Macmillan, Dom Massaro, Dick Pastore, David Pisoni, Philip Quinlan, Mark Steyvers, Michael Studdert-Kennedy, and Michel Treisman on an earlier draft of this paper enabled us to improve it markedly. The VOT stimuli used here were produced at Haskins Laboratories, New Haven, Connecticut, with assistance from NICHD Contract NOI-HD-5-2910. Thanks to Doug Whalen for his time and patience. Correspondence concerning this article should be addressed to R. I. Damper, Image, Speech and Intelligent Systems (ISIS) Research Group, Bldg. I, Department of Electronics and Computer Science, University of Southampton, Southampton S017 IBJ, England (e-mail: rid@ecs.soton.a...

A unified account of categorical effects in phonetic perception

“…Other neural network models have also been proposed to show a biologically plausible mechanism by which these categorical perception effect could arise. Damper & Harnad (2000) trained both a Brain-State-In-A-Box (BSB) (following Anderson et al, 1977) and a back-propogation neural network model to show how categorical perception arises through spontaneous generation after training on two endpoint stimuli. They were able to produce typical categorical effects and reproduce the discrepancy between VOT boundaries between different places of articulation found in human participants.…”

“…For stop consonants, the Liberman et al (1967) model comes close to predicting discrimination based on the identification function, while this prediction fails in vowel perception experiments. Neural network models have been applied to either vowel and liquid perception (Guenther and Gjaja, 1996; or to stop consonant perception (Damper & Harnad, 2000), but the same models have not typically been used to account for perception of both classes. While we now know that no phonemes are perceived purely categorically, the literature has nevertheless continued to treat strongly categorical perception as a separate phenomena from more continuous perception of other sounds, particularly vowels (Table 1).…”

“…Previous models of categorical perception have assumed that discrimination tasks primarily involved listeners' inferences of category labels (e.g., (Liberman et al, 1957), (Damper & Harnad, 2000)). Our model instead posits that listeners are primarily focused on recovering continuous phonetic detail in discrimination tasks, and recruit their knowledge of categories only because it helps them solve this inference problem.…”

The self-organization of speech sounds

“…As explained in (Oudeyer, 2001a), this system contains neural maps which are similar to those used in (Guenther and Gjaja, 1996) and (Damper and Harnad, 2000). What is different is that on the one hand, motor and perceptual neural maps are coupled so that the learning of sounds affects the production of sounds, and on the other hand, these two other works used single agents that learnt an existing sound system, while here we use several agents that co-create a sound system.…”

Modeling learned categorical perception in human vision

“…While there are successful neural models of category learning (Gluck & Bower, 1988;Kruschke, 1992;Jäkel, Schölkopf, & Wichmann, 2008), one question that arises from selecting a neural model is whether it will be sufficient to allow a CP effect to be generated? Damper & Harnad (2000) conducted a systematic review and comparative experiments to determine if neural models are valid in such studies. They concluded that if categorization is in some way built into a sensory stimulus, "any general learning system operating on broadly neural principles ought to exhibit the essentials of CP" (p.862).…”