Can computers learn to distinguish style from content like humans do? This research presents a computational model for separating 'content' from 'style' in perceptual systems using bilinear models. Perceptual systems routinely separate “content” from “style,” classifying familiar words spoken in an unfamiliar accent, identifying a font or handwriting style across letters, or recognizing a familiar face or object seen under unfamiliar viewing conditions. The general framework solves two-factor tasks using bilinear models and can be fit to data using efficient algorithms based on singular value decomposition and expectation-maximization. It provides expressive representations of factor interactions while maintaining computational tractability. Existing factor models (Mardia, Kent, & Bibby, 1979; Hinton & Zemel, 1994; Ghahramani, 1995; Bell & Sejnowski, 1995; Hinton, Dayan, Frey, & Neal, 1995; Dayan, Hinton, Neal, & Zemel, 1995; Hinton & Ghahramani, 1997) are either insufficiently rich to capture the complex interactions of perceptually meaningful factors such as phoneme and speaker accent or letter and font, or do not allow efficient learning algorithms. The model is tested across three perceptual domains: spoken vowel classification, font extrapolation, and face illumination translation. The model offers a powerful tool for machine learning and artificial intelligence, with potential applications ranging from speech recognition to image processing.
Published in Neural Computation, this paper aligns with the journal's focus on theoretical and computational approaches to understanding neural and cognitive processes. The development and application of bilinear models for separating style and content contribute to the understanding of perceptual learning and representation.