Ackley, Hinton & Sejnowski, A Learning Algorithm for Boltzmann Machines, 1985 · §4, fig. 2–4

Hinton DiagramsThe recursive weight matrix of a Boltzmann machine

Recursive diagram rows of boxes: V1, H, V2 · inside each box, the same layout again

White square: positive weight. Black square: negative. Area ∝ |w|, so side ∝ √|w|, scaled to the largest weight in the network. A dashed cell is the unit's own place: its bias (−θ). Every link appears twice, once in each of its two units.

Weight distribution every link once, and the biases

■ lateral links inside V1 and V2 ■ links between visible and hidden units ■ biases

Hidden codes most likely H for each symbol, at T = 10

Edit the weight matrix

JSON with W, an n × n array (rows and columns in the order of units), and b, the n biases. The matrix is made symmetric, wij = wji, and pairs that the encoder does not link (V1–V2, H–H) are set to 0.