Exploring Human and Neural Attention on Source Code: Insights and Applications

In the rapidly evolving landscape of AI-driven software engineering, understanding how neural models perceive code has become paramount. This talk delves into the fascinating commonalities and differences between human and neural attention in code-related tasks. In particular, we compare the reasoning of skilled developers with the attention mechanisms of neural models, including even recent Large Language Models (LLM), on tasks like code summarization, bug fixing and sense-making.

The results uncover correlations and divergences, shedding light on the potential and challenges of leveraging neural attention. We conclude by introducing the novel concept of follow-up attention that leverages the attention signal of LLMs to harness their knowledge for supporting developers in code exploration tasks.

Speaker – Matteo Paltenghi
Matteo Paltenghi is a doctoral researcher from University of Stuttgart with expertise at the intersection of artificial intelligence and software engineering. With a recent collaboration at GitHub Next, he harnessed Large Language Models for code exploration. His Ph.D. work, advised by Prof. Dr. sc. Michael Pradel, encompasses software engineering, AI, and quantum computing, with presentations at top conferences like ASE 21, OOPSLA 22 and ICSE 23. Prior to this, he spend 9 months at CERN for his Master Thesis on anomaly detection on data center.

Matteo holds a double degree M.Sc. Computer Science and Engineering from Politecnico di Milano and TU Berlin, proceeded by a B.Sc. from Politecnico di Milano. He recently started serving as reviewer (TOSEM, JSSoftware) and as session chair at MSR 23. Recently, he was also one of the few young researchers selected for participation in the Heidelberg Laureate Forum (HLF 23).

Meetup group – https://www.meetup.com/machine-learning-methods-in-software-engineering/

Видео Exploring Human and Neural Attention on Source Code: Insights and Applications канала JetBrains Research