Win Function

Win Function Norbert Ropiak

Figure 1 · how it works

Figure 1. The name above is the minimum of a loss function. Each of the 7 000 points is a parameter vector θ, started at random and moved by noisy gradient descent (Langevin dynamics). The loss starts as a random non-convex landscape and is annealed into one whose minimum spells the company’s name. The grey lines are its contours, and each point is coloured by its loss on the viridis scale. Move the pointer over the plot to raise a ridge; click it to train again.