Artificial intelligence; convolutional neural networks; neuroscience; geometric deep learning
Abstract :
[en] Some shapes look different to us if rotated. That is attributed to the use of a rotation frame of coordinates in the human visual system. However, no evidence that ConvNets, which is a machine learning architecture, use a frame of coordinates for rotation. We investigated the effect of adding one to ConvNets. An explicit orientation encoding kernel was developed using a mathematically inspired self-supervised approach. The experimental results showed that rotation encoding improved the accuracy of classifying rotated images and the resilience against noise. The orientation encoding kernel was embedded in a layer named (ConvOrient) that will be open-sourced.