Artificial Intelligence that Learn to Write Code : Memory Guided Programming

EasyChair Preprint 1579

Abstract

Writing code requires the brain to understand the meaning of language and to properly organize the thoughts flow using the language. However, current models to write its own working code are primarily limited to searching through a huge code database developed by various experienced programmers. Here, we proposed a Memory Guided Programming( MGP ) network to incrementally learn the meaning and usage of diverse functions / parameters, aiming to make best possible arrangement for a human-like machine programming process. MGP contains three subsystems : (1) Code system which consists of a mapping to transfer symbol texts into respective numeric and a RNN to generate the sequence dependencies from the input texts, and a output encoder to convert numeric values into text symbols; (2) Image system that contains an encoder to convert the input into abstract representations, and a DNN to classify image scenarios from real level representations; (3) a LSTM that combines inputs in the forms of both code and image, and predict text symbols and next images accordingly. In this work, the proposed MGP network illustrates the ability to incrementally learn different programming scripts and form a machine programming loop that enables interactions between Code and Image system. The paper presents an architecture that allows the machine to learn, understand and use programming language in a human-like way, which might enable a machine to construct programming scenarios and possibly possess human-like intelligence. 

Keyphrases: Artificial Intelligence, Human-like Programming, Machine Programming, Predictive Programming, Writing Own Code

@booklet{EasyChair:1579,
  author    = {Poondru Prithvinath Reddy},
  title     = {Artificial Intelligence that Learn to Write Code : Memory Guided Programming},
  howpublished = {EasyChair Preprint 1579},
  year      = {EasyChair, 2019}}