Figure 10.

Results of the Procrustes analysis for the Res architecture trained on the CC dataset (black) and with random weights (gray). The analysis was performed on the outcomes of the multidimensional scaling at different layers using Munsell space as reference coordinates. The variance explained for ResCC was consistently lower than for DeepCC throughout its layers, meaning that ResCC discriminate Munsells following color dimensions dissimilar to those defined by human color perception. The fact that the Res architectures systematically rate lower both when trained and with random weights suggests that the major factor for this difference is the architecture.