Summary
International Workshop on Smart Info-Media Systems in Asia
2021
Session Number:RS2
Session:
Number:RS2-7
Machine Code Generation for the RISC-V Instruction Set Using Neural Programmer-Interpreters
Masahiko Tsuyama, Ryusuke Miyamoto,
pp.83-88
Publication Date:2021/9/20
Online ISSN:2188-5079
PDF download
Summary:
Machine learning has become a popular method to solve tasks as the classification performance has been greatly improved since the emergence of deep neural networks. Neural Programmer-Interpreters (NPI) can learn algorithms when sample program sequences are given as training input. The NPI model is provided with the scratch pad that works as an external memory of the model. By updating the scratch pad, the NPI solves problems. It is trained with the example sequences of task execution consisting of subprograms that are the basic units of the program. The immediate subprograms of the subprograms are used to update the scratch pad. The most significant feature of the NPI is that it can learn how to solve a task from sample input sequences. This function makes it possible to generate a machine code running on target architecture while using a sample program running on another architecture, to generate training examples. To confirm this idea, this paper trained the NPI model targeting soft-multiplication with the training dataset consisting of RISC-V instruction set to generate machine code. The results of the evaluation show that our proposed method achieved 100 % accuracy.