Algorithmic transformations for neural computing and performance of supervised learning on a dataflow machine

摘要

Reprogrammable dataflow neural classifiers are proposed as an alternative to traditional implementations. In general, these classifiers are based on functional languages, neural-dataflow transformations, dataflow algorithmic transformations, and dataflow multiprocessors. An experimental approach is used to investigate the performance of a large-scale fine-grained dataflow classifier architecture. In this study, the functional descriptions of high level data dependency of a supervised learning algorithm are transformed into a machine executable low-level dataflow graph. The tagged token dataflow algorithmic transformation is applied to exploit the parallelism. Dataflow neural classifiers are used to implement the learning algorithm. No attempt is made to optimize the granularity of the high-level language programming blocks to balance the computation and communication. The proposed classifier architecture is more versatile than other existing architectures. Performance results show the effectiveness of dataflow neural classifiers.