notesum.ai

Published at December 4

A Granger-Causal Perspective on Gradient Descent with Application to Pruning

cs.LG

Released Date: December 4, 2024

Authors: Aditya Shah¹, Aditya Challa², Sravan Danda², Archana Mathur³, Snehanshu Saha²

Aff.: ¹Google Search, Google Austin, Texas, USA; ²APPCAIR and CS&IS, BITS Pilani KK Birla Goa Campus, India; ³Nitte Meenakshi Institute of Technology, Yelahanka, Bangalore, India

Arxiv: http://arxiv.org/pdf/2412.03035v1

Name	Explanation	Typical Values
$N_{pre}$	Number of epochs for training the network after which the pruning is performed	5-10
$N_{iter}$	Number of iterations of pruning to be performed	2-10
$N_{prune}$	Number of epochs for training the network to collect the data required for causal pruning	2-10
$N_{post}$	Number of epochs for training the network after pruning to evaluate the performance	100-300
L1_coeff	The regularization parameter used for LassoRegression	$110-18$ to $110-11$ (log space)