Conference on Cognitive Computational Neuroscience

Tremendous progress has been made in proposing models that try to explain image understanding in the human brain. However, the available models either lack high prediction accuracy for all visual areas or difficult to interpret with respect to the human visual system. To address this problem, the VOneCAE architecture is introduced in this paper. The VOneCAE model consists of two components: the VOne block which is aimed to improve the biological interpretability, and the Convolutional AutoEncoder (CAE) block, which is planned to construct a compressed feature space using unsupervised learning. Experiments reveal that the VOne block accurately predicts the early visual areas, such as V1 and V2, and the CAE block performs well for the late visual areas, such as V4 and IT. More precisely, the combination of two blocks performs well for all visual areas.