Controlling Recurrent Neural Networks by Diagonal Conceptors

TL;DR

This paper introduces diagonal conceptors, a simplified and computationally efficient variation of traditional conceptors for controlling recurrent neural networks, maintaining accuracy while reducing complexity.

Contribution

The paper proposes diagonal conceptors, a novel variation that simplifies control of RNNs by using diagonal matrices, making the approach more practical.

Findings

Diagonal conceptors achieve similar accuracy to full matrix conceptors.

They are computationally more efficient due to reduced matrix size.

Diagonal conceptors are slightly more unstable but remain promising.

Abstract

The human brain is capable of learning, memorizing, and regenerating a panoply of temporal patterns. A neuro-dynamical mechanism called conceptors offers a method for controlling the dynamics of a recurrent neural network by which a variety of temporal patterns can be learned and recalled. However, conceptors are matrices whose size scales quadratically with the number of neurons in the recurrent neural network, hence they quickly become impractical. In the work reported in this thesis, a variation of conceptors is introduced, called diagonal conceptors, which are diagonal matrices, thus reducing the computational cost drastically. It will be shown that diagonal conceptors achieve the same accuracy as conceptors, but are slightly more unstable. This instability can be improved, but requires further research. Nevertheless, diagonal conceptors show to be a promising practical alternative to the standard full matrix conceptors.