SGER: Learning Invariance for Recognition -- A Computational Approach

Trustees Of Boston University, Boston

Investigators

Abstract

PI: VAINA Abstract The objective of this proposal is to investigate computationally a biologically plausible model for learning invariance. Invariant recognition is a fundamental capacity of perceptual systems, which makes it possible to recognize visual objects under different viewing conditions, such as changes in the relative position of the object in the visual field, changes in distance, viewing direction, illumination, and also under shape deformations. The purpose of this study is to extend and generalize the shift invariance algorithm and incorporate a broader class of invariances, such as size and rotation in depth and combine it with a learning algorithm to achieve solutions to novel types of computational problems in biological vision. A major focus of this research will by on computational studies of biological visual problems for which there is ample neurophysiological and psychophysical data. Two important and novel characteristics of the models to be developed are: 1) invariance is achieved gradually in a series of processing stages, and 2) a simple unsupervised learning mechanism is sufficient for connecting units in a neural network in a way that results in invariant recognition. This approach is broader than previous approaches and therefore riskier. However, if it is shown to be correct, then it will offer a uniform account of multiple aspects of invariant perception and allow the development of powerful learnable mechanisms for computer vision systems.