Ilona Kulikovskikh

Ilona Kulikovskikh

Recent studies revealed complex convergence dynamics in gradient-based methods, which has been little understood so far. Changing the step size to balance between high convergence rate and small generalization error may not be sufficient: maximizing the test accuracy usually requires a larger learning rate than minimizing the training loss. To explore the dynamic bounds of convergence rate, this study introduces \textit{differential capability} into an optimization process, which measures whether the test accuracy increases as fast as a model approaches the decision boundary in a classification problem. The convergence analysis showed that: 1) a higher convergence rate leads to slower capability growth; 2) a lower convergence rate results in faster capability growth and decay; 3) regulating a convergence rate in either direction reduces differential capability.

Ilona Kulikovskikh, Tarzan Legović

Convergence and generalization are two crucial aspects of performance in neural networks. When analyzed separately, these properties may lead to contradictory results. Optimizing a convergence rate yields fast training, but does not guarantee the best generalization error. To avoid the conflict, recent studies suggest adopting a moderately large step size for optimizers, but the added value on the performance remains unclear. We propose the LIGHT function with the four configurations which regulate explicitly an improvement in convergence and generalization on testing. This contribution allows to: 1) improve both convergence and generalization of neural networks with no need to guarantee their stability; 2) build more reliable and explainable network architectures with no need for overparameterization. We refer to it as "painless" step size adaptation.

Ilona Kulikovskikh, Tarzan Legović

Current expectations from training deep learning models with gradient-based methods include: 1) transparency; 2) high convergence rates; 3) high inductive biases. While the state-of-art methods with adaptive learning rate schedules are fast, they still fail to meet the other two requirements. We suggest reconsidering neural network models in terms of single-species population dynamics where adaptation comes naturally from open-ended processes of "growth" and "harvesting". We show that the stochastic gradient descent (SGD) with two balanced pre-defined values of per capita growth and harvesting rates outperform the most common adaptive gradient methods in all of the three requirements.