基于类别结构学习启发式方法的自适应测试时计算分配

Deep neural networks often face significant computational constraints during deployment, particularly on edge devices. Existing approaches, such as static pruning or dynamic early exit strategies, typically suffer from limitations in the granularity of compute allocation and fail to fully leverage the rich categorical structure information embedded within the model. This paper introduces a novel adaptive test-time compute allocation framework that dynamically adjusts computational effort by learning heuristic policies. The core idea is to exploit the categorical structure information present in intermediate layers of the model, such as the separability of different classes in the feature space, to guide resource allocation. Specifically, we design a lightweight policy network that takes intermediate features and categorical structure metrics as input, predicting whether subsequent layers need to be executed. Through reinforcement learning or end-to-end training, this policy network learns decision rules that minimize computational cost while maintaining performance. Experimental results demonstrate that our method achieves substantial computational efficiency improvements across various benchmark datasets and models, while preserving accuracy comparable to the full model. It outperforms traditional static pruning and dynamic early exit methods. This framework offers a new paradigm for deploying deep learning models in resource-constrained environments, holding broad application prospects.