Path branching and Top-k sparsification achieve the segregation of specialized features

Standard deep neural networks rely heavily on texture rather than shape for object recognition, unlike human vision. While introducing activation sparsity enhances shape bias, sparse networks often suffer from accuracy degradation. Inspired by biological visual processing, we propose combining multi-path branching with asymmetric Top-k sparsification. Evaluated on ImageNet using a three-branch AlexNet architecture, applying Top-k sparsity to a single branch induces clear feature segregation: non-sparse branches specialize in grayscale Gabor filters and color patches, whereas the sparse branch isolates edge-detection filters. Fine-tuning on the sparse branch achieves an average shape bias of 0.85, comparable to human perception, while maintaining classification accuracy. These findings highlight signal-sparsity symmetry breaking as a vital mechanism for acquiring human-like neural representations.