Advanced Computing Trends in Latin America

Abstract

Deep neural networks (DNNs) achieve state-of-the-art performance in many areas, including computer vision, natural language processing, system configuration, and question-answering. However, DNNs are expensive to develop, both in intellectual effort (e.g., devising new architectures) and computational costs (e.g., training). Reusing DNNs offers a promising solution to amortize these costs, both within organizations and across the broader computing industry. As the paradigm shifts towards Software 2.0, however, understanding reuse patterns becomes critical. Trust in deep learning models is essential, particularly as they become integrated with traditional (non-ML) software. Software engineering must also evolve to account for the growing use of DNN components, including generated artifacts. In the era of generative AI, where models and prompts serve as valuable artifacts, preserving them is crucial for ensuring reproducibility. This invited talk describes the challenges in current DNN reuse, including both technical gaps and missing engineering practices. We summarize failures in reuse techniques across conceptual (e.g., reuse based on research), adaptive (e.g., building on existing implementations), and deployment (e.g., reuse on new devices) approaches, while outlining advances needed to improve each.