GRAD -CM2: A Data-Parallel Connectionist Network Simulator
Thomas A. Fontaine · ScholarlyCommons (University of Pennsylvania) · 1992
A data-parallel simulator capable of training recurrent time-delay connectionist networks is described. The simulator, GRAD-CM2, is written in the C* programming language and runs on a Connection Machine CM-2. GRAD-CM2 is an extension of the serial simulator, GRADSIM [8], and offers similar features. Timing performances of GRAD-CM2 and GRADSIM on a series of spatiotemporal discrimination tasks are presented to emphasize the efficacy of data-parallelism. 1 Introduction Although the realization that the error gradient of a multi-layer network could be computed (popularized by [4]) led to the application of a variety of powerful gradient minimization methods [7], the training of connectionist models in certain domains remains painstakingly slow. As the scope and complexity of connectionist applications expand, there is an increased demand for faster training methods. The time required to train a network is dependent upon several factors including the algorithm used, the network to...