Memorization Learning for Object Recognition

We consider the fundamental complexity of intelligent systems, and argue that some form of learning is essential in order to acquire the amount of information necessary to specify such a system. In this view, learning is essentially a means of acquiring large amounts of structured information by relatively simple recoding of information owing down high-bandwidth channels (the senses) from a pre-existing source of structured information (the world). This chapter considers the case of visual learning. We argue that, of the known methods of computational learning, the only technique that can eeectively assemble the very large amount of information needed to specify visual intelligence amounts to the construction of an indexed memory. We show how an associative memory organization coupled with the use of robust, semi-invariant features can lead to reliable 3-D object recognition. There has been an ongoing endeavor, at least since the industrial revolution, to construct machines that can do some of what people can do, only faster, better, and more cheaply. In terms of physical processes, this eeort has been enormously successful, and the resultant reduction in the amount of human labor required to produce physical artifacts has been largely responsible for the increase in the so-called standard of living that has occurred in the last two centuries. There is, however, a group of tasks that require abilities that are loosely thought of as intelligence, e.g., reasoning or sophisticated sensing. Performing such tasks seems to involve, fundamentally, a complexity that is not practical to implement in a macroscopic mechanical device. With the advent of digital computers, however, came the idea that the required complexity might be reducible to information processing operations. For performing such operations, electronic devices can be built that are many orders of magnitude faster, more compact, more reliable, and less expensive than macroscopic mechanical devices, and it is still not clear 1