The effects of varying memory vector size in a network that learns to learn

Simulation results show that DARNET, a network model that learns using a gradient-descent procedure to perform single-trial learning, can make efficient use of whatever number of memory trace vector elements it is provided with. However, the effective maximum capacity of the system is determined by the architecture of the model to be the number of items that can be stored in a composite memory trace vector of fixed dimensionality, the optimal size depending on the number of elements in each to-be-associated input vector. DARNET has previously been shown to provide a good account of some relevant psychological data and the present work adds to the authors' understanding of the constraints governing its memory capacity and ability to "learn-to-learn".<
>