Abstract
The increasing adoption of Artificial Intelligence (AI) necessitates the deployment of Deep Neural Networks across various computational environments. This paper addresses the dynamic nature of computational requirements and poses the question: How can we adjust the complexity of an AI classification system to maximize accuracy while adhering to changing computational constraints? We refer to this challenge as Budgeted Image Classification and formally frame it as a resource allocation integer program.
Given a computational budget, a batch of images, and a classification system with multiple decision points, we investigate strategies to allocate images to these decision points to maximize accuracy within the available budget. The original integer program is NP-Hard; hence, we propose a continuous relaxation leading to a content-agnostic allocation strategy that assigns images to decision points without regard to their specific content. To address this limitation, we introduce a content-sensitive strategy, which we experimentally demonstrate to yield superior performance.
We theoretically analyze the behavior of our strategies, deriving necessary conditions for decision points to be suitable for budgeted classification. Additionally, we examine failure cases and provide insights for future research directions.
Blogger's Review: This paper effectively tackles the complexity issue in budgeted image classification through dynamic resource allocation, showcasing strong practical value. The introduction of a content-sensitive strategy highlights the necessity of considering content in AI decision-making, paving the way for further exploration in various application scenarios.