Skip to main content
. 2022 Apr 14;16:836498. doi: 10.3389/fncom.2022.836498

Figure 1.

Figure 1

Different approaches to dealing with symmetries in neural networks. Both figures represent a neural network transforming an image, and a rotated image. The gray 3 × 3 squares are activations of the neural networks. (A) Inputs transform with symmetries, but hidden features and outputs are invariant. (B) Inputs and hidden features transform with symmetries, only outputs are invariant.