6d3a2d24eb109dddf78374fe5d0ee067-AuthorFeedback.pdf

This property is induced by our sparsity regularizer (2) that6 minimizestheL0 normoftheadjacencymatrix. Isomap constructs knn-or -graph, then computes shortest paths in that graph and17 approximates these paths with Multi-Dimensional Scaling in Euclidean space. We verify the robustness of our training procedure by28 running several experiments with different random initializations and different initial numbers ofneighbors. Figure29 1 shows the learning curves of PRODIGE under various conditions.