PERTINENCE: Input-based Opportunistic Neural Network Dynamic Execution

A runtime method called PERTINENCE can dynamically select lighter neural network models for simpler inputs, reducing computational operations by up to 36% while matching or improving accuracy, according to a paper posted on arXiv [1]. Deep neural networks are widely used for modeling complex patterns in computer vision, speech recognition, and robotics, but larger models are computationally expensive and energy-intensive [2]. The cost is typically necessary only for challenging inputs, so dynamically selecting lighter models for simpler inputs can improve efficiency with minimal impact on accuracy [2]. Marcello Traiola and colleagues introduce PERTINENCE, a runtime method that selects, from a set of pre-trained models, the lightest model likely to process each input correctly [1]. An ML-based dispatcher performs this selection, and a genetic algorithm explores dispatcher training strategies to identify Pareto-optimal trade-offs between accuracy and computational cost [1]. The researchers evaluated PERTINENCE on convolutional neural networks trained on CIFAR-10 and CIFAR-100, vision transformers trained on TinyImageNet, and a YOLO-based road occupancy estimation application using real-time intersection camera feeds [1]. Results show that PERTINENCE matches or improves the accuracy of state-of-the-art pre-trained models while reducing operations by up to 36%, with equivalent or lower end-to-end inference time through tunable invocation intervals [1]. The paper was submitted to arXiv on July 2, 2025, and revised most recently on June 24, 2026 [1]. arXiv is an open-access repository of electronic preprints and postprints approved for posting after moderation but not peer reviewed, covering fields including computer science and statistics [6]. As of November 2024, the submission rate to arXiv was about 24,000 articles per month [6]. The repository passed the two-million-article milestone by the end of 2021 [6]. The PERTINENCE paper appears on arXiv’s abstract page alongside experimental community tools developed through arXivLabs, a framework that allows collaborators to develop and share new features directly on the site [4]. arXivLabs projects include the Bibliographic Explorer, which displays citation information, and the CORE Recommender, which facilitates exploration of relevant open access papers from a global network of research repositories [5]. arXiv states that third-party collaborators have access only to minimal and anonymized user data, strictly for ensuring correct functioning of the Labs features [4].