Impact of a biomimetic training regimen based on early visual experience on neural network organization and behavior

While deep convolutional neural networks have emerged as powerful computational model systems, they are known to be primarily driven by local features, rather than global shape information, and typically lack robustness to stimulus perturbations that degrade input quality. In this study, we examine the impact of incorporating insights from human development into the training of deep neural networks for vision. Specifically, we focus on the role of experience with initially degraded sensory inputs characteristic of early visual development. Previous work demonstrated that commencing deep network training with initially degraded inputs along isolated perceptual dimensions (specifically, visual blur and color degradations) improved generalization performance. Here, inspired by the joint developmental trajectories of newborns, we examined the consequences of ‘biomimetic’ training regimens that transitioned from blurry, achromatic to non-blurry, full-color visual inputs. These simulations reveal that such biomimetic training induces a more human-like bias to derive classification decisions based on global shape, rather than local texture. Further, receptive field analyses suggest that the joint development of spatial frequency and chromatic sensitivities can provide a candidate account for the emergence of the division of the visual pathway into parvo- and magnocellular systems. Ablation studies further suggest that magnocellular-like receptive fields are causally driving the shape bias of the biomimetic network. These results have important implications for understanding a key aspect of visual pathway organization and hold applied significance for enhancing deep learning training procedures based on incorporating developmental aspects of human behavior.