Addressing a Problem in Constructing a Transcoder using Neural Programmer-Interpreters
-
- Masahiko Tsuyama
- Meiji University
-
- Ryusuke Miyamoto
- Meiji University
抄録
Recent advances in deep neural networks have improved performance of several types of tasks, although some of them cannot be modeled by machine learning schemes. Neural Programmer-Interpreters (NPI) can learn algorithm when exam- ple traces of the task are given as the training input. Our previous study proposed a machine code generation scheme using NPI: scratch-pad, subprograms, and immediate-subprograms were modified to treat instructions included in RISC-V architecture appropriately. Experimental results showed that the NPI model trained using several examples that executes soft-multiplication can learn methods to perform it using RISC-V instruction sets; the accuracy reached 100% in the previous research. This study attempts to construct a transcoder that generates machine code for an architecture using training data about a task represented with machine codes for another architecture. However, there are some technical limitations to the application of the NPI model to construct a transcoder. To address these problems, this study proposed a novel training scheme for the NPI model and tried to generate programs composed of RISC-V instructions from the NPI model trained using machine code for x86 64 instructions. Evaluating the trained model by diagnosing the output sequence from the NPI model, the RISC-V instructions were correctly replaced with x86 64 instructions. However, the arguments were not completely generated.
収録刊行物
-
- IEICE Proceeding Series
-
IEICE Proceeding Series 69 57-62, 2022-09-15
The Institute of Electronics, Information and Communication Engineers
- Tweet
詳細情報 詳細情報について
-
- CRID
- 1390293943108430080
-
- ISSN
- 21885079
-
- 本文言語コード
- en
-
- データソース種別
-
- JaLC
-
- 抄録ライセンスフラグ
- 使用不可