LOCALLY LEARNED REGULARITY IN ARTIFICIAL NEURAL NETS

Abstract

We present our study on local camera adjustment with simple neurons, and local avoidance of chaos by neurons with more complex dynamics. These works are based on the notion of locally learned regularity in neural nets, i.e., neural nets learning ‘good’ behaviour by local, scalable means, rather than learning that slows exponentially with size increase. In the first part of the thesis, we give a neural net approach to correcting the image distortions produced by camera configurations, to produce a geometrically ‘ideal’ image. The network learns by comparison of actual and desired versions of response to standard inputs, and then produces corrected versions of response to standard inputs, and then produces corrected versions of arbitrary images. The correction system once learned is rapid, and easy to configure into a chip for the particular camera. Correcting a new camera requires only an automated learning process, not an engineering analysis of the sources of error. The second part examines simple mechanisms of synaptic plasticity that enable a chaotic neural unit to learn sets of prescribed stable outputs, in spite of the unit’s dynamics having both regular and chaotic behavioural regimes, as displayed experimentally by neurons.