Neural Models for Part - Whole

We present a connectionist method for representing images that explicitly addresses their hierarchical nature. It blends data from neu-roscience about whole-object viewpoint sensitive cells in inferotem-poral cortex 7 and attentional basis-eld modulation in V4 3 with ideas about hierarchical descriptions based on microfeatures. 4, 10 The resulting model makes critical use of pathways for both analysis and synthesis. 5 We illustrate the model with a simple example of representing information about faces. Substantial recent eeort has been devoted to analysis-by-synthesis models for visual processing. The synthetic or generative models form the map: `object' ! `image' (1) wherèobject' implies the identity of the object (such as a face) and its relevant instantiation parameters (such as whether it is smiling or frowning) andìmage' stands in for the pattern of activity over the input units that is caused by observing the object. The Helmholtz machine 5 (HM) employs unsupervised learning to construct an explicit generative model of this form, using a top-down layered processing structure that is a form of factorial belief network. 9 The activities of the units in the top layer, or possibly all hidden layers, represent the object and the choices of its instantiation parameters. The task for analysis or recognition is to take anìmage' and produce (a probability distribution over) the object that might have generated it, and also the instantiation parameters of that object. The HM uses a bottom-up belief net over the same units to represent the recognition model, and trains top-down and bottom-up models to be mutual inverses.