High Capacity RAAM

Author: Jordan Pollack (Brendeis University)

Abstract: RAAM pioneered the recursive use of auto-associators, allowing a neural network to devise its own representations for compositional data structures such as parse-trees and logical forms. However, because of a lack of understanding of the dynamics of iterated networks, the binary "terminal test" for RAAM caused limited capacity The symbolic dynamics of most recurrent networks, and of RAAM decoders, are related to fractals, we will demonstrate RAAM-2 which uses "on the attractor" as its terminal test, and is capable of much higher capacity to represent constitute structures as finite dimensional activation patterns.