Things to discuss (during that meeting or at another time):
How are the ICNs reconstructed from the feature maps in https://github.com/gkiar/reproreading/issues/44? This is important in their application and also relevant to JB's comment about incorporating constraints in the autoencoder based on the "usual" decomposition of functional maps.
Things to discuss (during that meeting or at another time):