What You See is What it Means! Semantic Representation Learning of Code based on Visualization and Transfer Learning

Recent successes in training word embeddings for Natural Language Processing ( NLP ) tasks have encouraged a wave of research on representation learning source code, which builds similar methods. The overall objective is then to produce code that capture the maximum program semantics. State-of-the-art approaches invariably rely syntactic (i.e., raw lexical tokens, abstract syntax trees, or intermediate tokens) generate embeddings, are criticized literature as non-robust non-generalizable. In this work, we investigate novel embedding approach based intuition has visual patterns We further use these address outstanding challenge identifying semantic clones. propose WySiWiM ‘ ‘What You See Is What It Means ” where representations fed into powerful pre-trained image classification neural networks from field computer vision benefit practical advantages transfer learning. evaluate proposed task vulnerable prediction and two variations clone identification: detection (a binary problem), multi-classification problem). show with experiments BigCloneBench (Java), Open Judge (C) although simple, our performs effectively state-of-the-art such ASTNN TBCNN. also showed data NVD SARD can be used learn detector reasonable performance (accuracy ∼90%). explore influence different steps approach, choice algorithm, eventually discuss promises limitations direction.