42 architectures and training schemes, from the perceptron (1958) to diffusion LLMs (2025).
Each diagram draws a network in miniature, from input at the bottom to output at the top. In most panels, a circle is one neuron; in sequence models, where each column is a token, it is that token's whole vector at one layer. A square is larger: an image patch, an expert network, a matrix-valued state. A cell's color and mark both show its role, so the diagrams read with or without color. Faded cells and links take little or no part in the step shown. Links show which tokens can read which. Wherever a part repeats, across positions, windows, copies, or passes, it uses the same weights. Each panel gives a year and the objective or learning rule of its original paper; many can be trained other ways. Select a diagram for its explanation, key, and papers.