Figure 2
From: Neuronal diversity can improve machine learning for physics and beyond
Schematic stochastic gradient descent meta-learning nested loops. Neural-network weights and biases \(\theta\) adjust to lower losses \({\mathscr {L}}(\theta ,\theta _s)\), during an inner loop, while periodically the sub-network weights \(\theta _s\) open extra dimensions and themselves adjust to allow even lower losses, during an outer loop. Rainbow colors code time t.