DARTS 2D: Architecture-Weight Arc Diagram

Neural architecture search normally means training thousands of candidate networks from scratch — DARTS makes it tractable by building one shared "supernet" where every edge is a continuous softmax mixture of every candidate operation, then using plain gradient ascent to push each edge's architecture weights toward whichever operation performs best. This 2D companion renders the identical mechanism as a flat arc diagram: six cell nodes sit on a line, each incoming edge arcs above it with a small composition bar showing its live 4-way softmax mixture, and a time-series panel beneath tracks architecture entropy, estimated validation score and training loss iteration by iteration. Adjust the architecture learning rate and softmax temperature, step or play the search, and watch the mixture bars sharpen from an even blend into one dominant operation per edge — then hit Discretize to prune the supernet down to the single derived architecture.