Human sensory systems exhibit higher sensitivity to common environmental features. This study investigates whether artificial neural networks trained on object recognition also demonstrate this efficient coding property. The authors mathematically show that gradient descent learning preferentially creates representations sensitive to common features, regardless of coding resource constraints and across supervised and unsupervised learning objectives. This finding suggests efficient codes are a natural outcome of gradient-like learning processes.

Ari S. Benjamin, Ling-Qi Zhang, Cheng Qiu, Alan A. Stocker, Konrad P. Kording