SOMO-VCB: A Matlab® software for single-objective and multi-objective optimization for variance counterbalancing in stochastic learning
Dimitra G. Triantali, Konstantinos E. Parsopoulos, I.E. Lagaris · Software Impacts · 2023
Variance CounterBalancing (VCB) is an effective neural network training method that outperforms traditional stochastic gradient descent techniques. By minimizing both the average and the variance of the network's error over random mini-batches, VCB improves generalization scores and solves the diminishing step-size issue. The SOMO-VCB software is a Matlab® implementation of VCB that can be used for regression problems as a single-objective or multi-objective optimization task. The code is easy to modify and flexible, using only essential Matlab® programming. Recent experiments show that SOMO-VCB is highly competitive compared to state-of-the-art methods like Adam.