Fig. 5

Learning to solve ≈10³⁷ “Hidden clique” problems. (Bottom) Adjacency matrices of three 64-cliques on \(v = 128\) vertices. (Top) Adjacency matrices of noisy versions of the cliques having, on average, 1219 bits corrupted out of \(n = 8128\) from the original. Converging dynamics of a symmetric 3-parameter network \((x, y, z) = (0.0107, 0, 1)\) with minimum probability flow initialized at these noisy cliques uncovers the originals