Synthesizing neural network applications using computer algebra

The software engineering of neural network (NN) applications faces many different problems from those of traditional programming. Due to the stochastic nature of NNs, it is extremely hard to verify the correctness of a NN application by exhaustive testing. It is also very hard to tell whether a NN application is generating the intended results, because there are many parameters and variables. Lastly, NN software is usually time intensive and highly optimised code is required. This paper explores the use of computer algebra to synthesize neural network applications. The author can generate highly reliable and efficient codes (probably the most efficient possible) from a high level NN algebraic specification. The computer algebra package not only handles most of the algebraic manipulation in the derivation of the computation formulas, but also performs simplification and verification. Due to the non-existence of control codes, synthesized NN codes can fully exploit computer architecture such as high-speed cache pipelines.