Budgeted Image Classification with Content-Sensitive Resource Allocation

A new arXiv preprint (2607.23997) introduces Budgeted Image Classification, a framework for dynamically adjusting deep neural network complexity under changing computational constraints. The problem is formulated as a resource allocation integer program: given a budget, a batch of images, and a multi-decision-point classifier, the goal is to allocate images to decision points to maximize accuracy. Since the integer program is NP-hard, the authors propose a continuous relaxation leading to a content-agnostic allocation strategy. The work addresses the growing need for AI deployment in dynamic environments where computational resources vary.