Fig. 3From: Unsupervised relational inference using masked reconstructionIllustration of GINA. Snapshots are processed independently. Each input/snapshot associates each node with a node state (blue, pink). The output is a distribution over states for each node. During training, this distribution is optimized w.r.t the input. The output is computed based on a multiplication with the current adjacency matrix candidate (stored as C) and the application of a node-wise MLP. Ultimately, we are interested in a binarized version of the adjacency matrix. Color/filling indicates the state, shape identifies nodesBack to article page