Extended Data Fig. 4: Investigation and ablation studies of the design principles of AdaptiveNN.

All the results are reported on ImageNet. See Supplementary Section 3 for the details of comparative baselines. a, Efficacy of different methodologies for establishing the fixation localization strategy within AdaptiveNN. For a clean comparison, we train a classifier using only the features from visual fixations, and assume all samples use the same number of fixations, such that the resulting validation accuracy serves as a well-controlled measure to assess the effectiveness of each variant. Moreover, we consider an extensive variety of baselines for comparison, including selecting fixations using i) pre-defined rules; ii) goal-directed importance maps obtained by CAM (class activation map) algorithms; iii) CAM augmented with a Gaussian mixture model (GMM); and iv) policy networks learned using other algorithms. b, Average test loss corresponding to the validation data with different state values predicted by the Vision Agent in AdaptiveNN. We examine the state values taken from every step of sequential perception processes. c, Comparisons of different termination criteria for concluding the sequential perception process of AdaptiveNN. The term ‘anti-’ refers to the inverse of our proposed method (detailed in Supplementary Section 4.1), namely terminating the observation process for samples with relatively higher state values. Exact P ( > 0.0001) values: 0.00018, 0.00047. d–e, Comparisons with representative methodologies designed to improve deep learning models’ computational efficiency. Specifically, d evaluates against baselines that leverage the spatial redundancy or sample-wise redundancy in visual data. e examines models with multi-exit architectures (using the same backbones as AdaptiveNN) that allow for online computational cost adjustments. Exact P ( > 0.0001) value in (e): 0.00034. In a–c, the results show means ± standard deviations from five independent trials with different random seeds. *Two-sided independent samples t-test.