Poggio’s lab is currently focusing on a type of object recognition called immediate recognition. This phenomenon was first described in 1969 in a paper by MIT lecturer Mary Potter (now a professor of psychology at BCS) and her research assistant, Ellen Levy. Immediate recognition is the fastest known form of recognition. A subject in a classic immediate-recognition experiment is seated before a display and asked to push one of two keys in response to each image in a series, depending on whether it contains an animal or not. To make sure looking at one image doesn’t accidentally help subjects learn how to look at others, researchers choose images that are very different: many species, in many different poses and perspectives, set against a wide range of backgrounds. The photos come and go in a few tenths of a second. At the start of a study, a subject might have next to no awareness of even being shown an image, let alone recognizing what is in it. Yet amazingly, people hit the right keys more often than not. They get steadily better – and become conscious of the appearance of the images – with practice. Still, at the outset, something in the brain is able to recognize and categorize objects before the subject is even aware of seeing anything.
Immediate recognition is important to researchers because it is the simplest possible case of general object recognition. It happens too quickly to involve recruiting lots of neurons or processing information intensively or sending and receiving impulses over more than a fraction of a centimeter. Information from eye movements, a key element in other kinds of recognition (as in DiCarlo’s work), can play no role. Yet somehow the right keys get pressed (mostly), which means that a limited form of general-purpose object recognition must be possible using a relatively small number of neurons organized in a relatively simple fashion.
Building on work Poggio did with Max Riesenhuber, PhD ‘00, then a grad student at MIT and now a professor at Georgetown University, Serre, Poggio, and others in Poggio’s
group developed a theory about the part of the visual cortex chiefly responsible for immediate recognition. Their approach to visual processing was in many respects different from a machine vision engineer’s. For instance, most machine vision programs feature one processor executing a series of instructions in consecutive order, an architecture known as “serial processing.” The brain, on the other hand, uses “parallel processing,” an approach in which a problem is broken up into many pieces, each tackled separately by its own processor, after which the results are combined or integrated to get a single general result – say, the perception of a cow. In theory, engineers could use parallel processing for machine vision programs (and some have tried), but in practice it is seldom obvious how to break down a problem in a way that allows the finished pieces to be seamlessly recombined.