So here ParticleNet + QCDorWhat =Graph Autoencoder To do this, require: An update step already described Something to make a list of 4 vectors into a graph use a topK algorithm connect a 4 vector with its K nearest neigbours something to reduce the number of nodes something to increase the number of nodes afterwards again similar to a Pooling operation for a convolutional network Seems simple enough but if you look at the literature slow...and the benefits...are less clear (arXiv:1907.09000) advance...has lagged behind (arXiv:1907.00481) one cannot simply pool ... (arXiv:1806.08804) project the graph on a learnable axis combine neigbour nodes on this axis relearn the graph or use a graph combination rule maybe instead of multiple images below each other doo this on multiple pages let each node grow into a learnable graph combine the new graphs with the existing one