Neural Networks Think Without Thinking: Empirical Validation and Theoretical Framework

This work presents an empirical proof and theoretical framework for the hierarchical statistical-pathway hypothesis in deep convolutional neural networks. Using ResNet-18 trained on CIFAR-10, we show that a small subset of convolutional filters concentrates most task-relevant activations, and ablating these filters catastrophically degrades performance. We complement these results with a rigorous mathematical analysis of sparse pathways in two-layer ReLU networks under convex training, proving finite-support solutions, stability under perturbations, and direct correspondence between activation dominance and geometric alignment. Experiments across architectures (ResNet-34, VGG-16, ViT) confirm the universality of the phenomenon. These findings establish a clear mechanism by which deep networks “think without thinking,” reusing a stable set of dominant filters to achieve robustness and generalization.