Stochastic Optimization with Heavy-Tailed Noise via Accelerated Gradient Clipping

Authors: Eduard Gorbunov, Marina Danilova, Alexander Gasnikov

NeurIPS 2020 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We conduct several numerical experiments with the proposed methods in order to justify the theory we develop. In particular, we show that clipped-SSTM can outperform SGD and clipped-SGD in practice even without using large batchsizes.
Researcher Affiliation Academia Eduard Gorbunov MIPT and HSE, Russia Marina Danilova ICS RAS and MIPT, Russia Alexander Gasnikov MIPT and HSE, Russia
Pseudocode Yes Algorithm 1 Clipped Stochastic Similar Triangles Method (clipped-SSTM)
Open Source Code Yes One can find the code here: https://github.com/eduardgorbunov/accelerated_clipping.
Open Datasets Yes We have tested4 clipped-SSTM and clipped-SGD on the logistic regression problem, the datasets were taken from LIBSVM library [4].
Dataset Splits No The paper mentions using datasets from LIBSVM library but does not explicitly state the training, validation, or test dataset splits.
Hardware Specification No The paper does not provide specific hardware details (e.g., GPU/CPU models, memory amounts) used for running its experiments.
Software Dependencies No To implement methods we use Python 3.7 and standard libraries.
Experiment Setup Yes For all methods we used constant batchsizes m, stepsizes and clipping levels were tuned, see Section H.2 for the details.