Networks That Learn for Image Understanding

Learning is becoming a central problem in trying to understand intelligence and in trying to develop intelligent machines. This paper describes some recent work on developing machines that learn in the domains of vision and graphics. We will introduce an underlying theory which connects function approximation techniques, neural network architectures and statistical methods. While these techniques have limitations, one can overcome these limitations by using the idea of virtual examples. We shall describe some learning-based systems we have developed that recognize objects, in particular faces, nd speciic objects in cluttered scenes, and produce novel images under user control. Finally, we will discuss about the implications of this research on how the brain might work.