Reusing Deep Learning Models: Challenges and Directions in Software Engineering

TL;DR

This vision paper addresses the problem of high development costs for deep neural networks by examining how reusing DNNs can amortize effort across organizations. It offers a taxonomy of three reuse paradigms (conceptual, adaptation, deployment), analyzes the challenges faced in each, and proposes concrete directions for artifacts, tooling, auditing, and supply-chain trust. By connecting reuse with foundational models and DL interoperability, the work highlights practical paths toward standardized engineering practices and trustworthy deployment. The findings aim to guide both researchers and practitioners in building scalable, cost-efficient, and reliable DNN reuse ecosystems.

Abstract

Deep neural networks (DNNs) achieve state-of-the-art performance in many areas, including computer vision, 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 is a promising direction to amortize costs within a company and across the computing industry. As with any new technology, however, there are many challenges in reusing DNNs. These challenges include both missing technical capabilities and missing engineering practices. This vision paper describes challenges in current approaches to DNN re-use. We summarize studies of re-use failures across the spectrum of re-use techniques, including conceptual (e.g., reusing based on a research paper), adaptation (e.g., re-using by building on an existing implementation), and deployment (e.g., direct re-use on a new device). We outline possible advances that would improve each kind of re-use.

Figures (8)

• Figure 1: Deep neural network reuse is the process of using existing DNN technology for another purpose. We focus on three distinct types: conceptual reuse, where existing theory is repurposed; adaptation reuse, where existing DNN models are modified; and deployment reuse where existing DNN models are converted for use in a new environment. Dashed boxes provide examples of each type.
• Figure 2: Components of a deep neural network (DNN), represented at different levels of abstraction. A DNN is a composition of weighted operations. These are combined into a layer; a group of layers into a block; and a group of blocks into a sub-graph such as a "backbone" or a "head".
• Figure 3: Illustration of a data pipeline following the Extract-Transform-Load design pattern. The specific pipeline is for the You-Only-Look-Once (YOLO) model family Redmon2018YOLO.
• Figure 4: Overview of the conceptual reuse of deep learning models. An engineer learns about deep learning ideas from existing implementations, e.g., research papers and prototypes. They use this to guide their own development of a deep learning model, e.g., implementing it directly, reproducing the model, or replicating the model.
• Figure 5: A pre-trained deep neural network model (PTM) can be adapted for different tasks via transfer learning, quantization and pruning, and knowledge distillation. This figure is reused from Jiang2022PTMSupplyChain with their permission.