"Using Top-Down Connections During Recognition to Implement Flexible Supervised Neural Networks"
-- Tsvi Achler, IBM Research
Abstract
Understanding how neural networks perform recognition, and the internal form in which recognition information is stored, is essential to understanding how the brain interacts with the world.
Unfortunately, neural models of recognition, such as feedforward PDP networks (McClelland & Rummelheart), have substantial problems quickly updating new information and recalling learned patterns. Moreover, evidence supporting these models remains murky. Feedforward models, including PDP networks, have an iterative learning phase and a simpler recognition phase.
My work focuses on the hypothesis that a negative feedback iterative process occurs during recognition instead of learning. This increases functionality and moves computational complexity from learning to recognition. Thus neural network learning becomes faster, more flexible, simpler and easier to recall while recognition becomes more complex.
The complexity occurring during recognition presents intriguing findings. Although in certain cases recognition might be slower, these are the same scenarios where the brain is slower -- inherently displaying cognitive phenomena.
