Internal Supervision & Clustering: A New Lesson from ‘Old’ Findings?

Abstract

Learning is most often treated as a psychologically rich process in the extant literature. In turn, this has a number of negative implications for clustering in machine learning (i.e. grouping a set of objects so that objects in the same group – cluster – resemble each other more than they resemble members of other groups) to the extent that psychologically rich processes are in principle harder to model. In this paper, I question the view of learning as psychologically rich and argue that mechanisms dedicated to perception and storage of information could also be used in categorization tasks. More specifically, I identify the minimum resources required for learning in the human mind, and argue that learning is greatly facilitated by top-down effects in perception. Modeling the processes responsible for these top-down effects would make modeling tasks like clustering simpler as well as more effective. For clustering is seen here as building upon associations between perceptual features, while connection weightings and top-down effects substitute external supervision in executing the function of error identification and rewards.