Основные понятия
Hybrid approach combining diagonal Hessian approximation and fuzzy logic for efficient optimization.
Аннотация
This research introduces a hybrid approach to accelerate convergence in second-order optimization by utilizing an online finite difference approximation of the diagonal Hessian matrix and fuzzy inferencing of hyperparameters. The content covers deep learning models, second-order optimization methods, various techniques like CG, Newton's method, quasi-Newton methods, stochastic quasi-Newton methods, HF method, sub-sampled HF method, and more. It delves into the diagonal Hessian approximation technique, finite differences, and methodologies proposed by researchers. Additionally, it discusses Fuzzy Logic Based Scheduling with a literature review on Fuzzy Expert Systems and their applications in learning rate optimization. The content also presents a new method for diagonal Hessian approximation in deep learning optimization. Experimental results on ImageNet dataset comparing SALO against Adam, AdamW, and SGD are discussed along with empirical performance analysis.
Structure:
Introduction
Second Order Optimization Review
Diagonal Hessian Approximation
Diagonal Approximation of the Hessian by Finite Differences for Unconstrained Optimization
Introducing Our Method: Diagonal Hessian Approximation
Fuzzy Logic Based Scheduling: Literature Review
Empirical Performance: Application on ImageNet
Conclusion and Future Work
Статистика
Competitive results have been achieved.
Initial results have shown substantial improvement in accuracy.
Computational complexity is reduced using diagonal Hessian approximation.
Finite differences help calculate necessary partial derivatives.
Various techniques like CG methods, Newton's method are discussed.
Цитаты
"An online finite difference approximation of the diagonal Hessian matrix will be introduced."
"Competitive results can be delivered with this hybrid approach."