Fig. 4: Gradient descent in materia for realizing benchmark tasks.
From: Gradient descent in materia through homodyne gradient extraction
a Output currents of the optimized Boolean logic gates at 77 K. The black dashed line separates the “high” and “low” logic states. The output signal was processed such that the ramping up and down between the different logic configurations (which takes 15 ms, including waiting time) is removed. b 3D scatter plot showing the position of the test data points for the sphere classification in the 3D feature space composed of the voltages of the 3 input electrodes. The colour map indicates the optimized current value corresponding to each point, after the HGE-based training. c Scatter plot of the output current values as a function of distance to the sphere centre (radius) for the test data points, after the training. The black dashed line represents the threshold between the 0 and 1 current levels of the two classes (orange and green). d Value of the loss function versus the number of iterations of the gradient descent algorithm.
