The model has three steps: feature extraction from LFFs using p -th order gradient filters; (optional) transformation into rotation-invariant features for embedding-agnostic representations (otherwise, representations are embedding-aware); and mapping features into latent space using a multilayer perceptron.
