Abstract

This chapter focuses on John Locke’s views on the faculty of perception. Locke proposed several forms of perceptual abstraction in his attempts to explain how we acquired general category representations from interaction with their examples, and this chapter illustrates how some of his ideas have been realized in deep convolutional neural networks, the kind of architecture behind AlexNet. It further explains how four forms of abstraction which have sometimes been considered competitors actually cooperate in deep convolutional neural networks, through iteration of convolution and pooling nodes, to create a more powerful form of learning which is dubbed “transformational abstraction.” It concludes by reviewing challenges to the idea that deep convolutional neural networks are on their own a road to artificial intelligence such as their sample inefficiency and apparent brittleness.