EE Seminar: Batch Estimators for Regression Problems
Electrical Engineering Systems Seminar
Speaker: Inbar Hasidim
M.Sc. student under the supervision of Prof. Ofer Shayevitz & Prof. Meir Feder
Sunday, 10th November 2024, at 15:00
Room 011, Kitot Building, Faculty of Engineering
Batch Estimators for Regression Problems
Abstract
In various machine-learning scenarios, algorithms that divide data into batches are widely used. Separating the data to batches is often used because of computational constraints and to improve generalization. A common technique of calculating an estimator using batch partitioning is to calculate the estimator for each batch and then merge them by simple averaging. This method collapses for batch sizes that are not linear with the number of samples. To address the problem, our research introduces two novel algorithms that combine the batch estimators using a different approach. We examine these batch partitioning algorithms within the context of an overparameterized linear regression model with isotropic Gaussian features. We present lower and upper bounds for one of the estimators and employ a series of extensive numerical experiments on both of them aimed at elucidating their performance characteristics and behavior across diverse scenarios.
השתתפות בסמינר תיתן קרדיט שמיעה = עפ"י רישום שם מלא + מספר ת.ז. בדף הנוכחות שיועבר באולם במהלך הסמינר
