HLGM: A Novel Methodology For Improving Model Accuracy Using Saliency-Guided High and Low Gradient Masking

Ali Karkehabadi, Banafsheh Saber Latibari, Houman Homayoun, Avesta Sasan

Abstract: In this paper, we introduce the High and Low Gradient Masking (HLGM) approach, a groundbreaking saliency-guided training method that effectively enhances both the accuracy and the quality of saliency maps in computer vision models. This method stands apart from traditional saliency-guided training, which often compromises accuracy. HLGM employs a novel two-phase process: initially, it involves regular training without gradient masking, followed by an accuracy boosting phase. This phase alternates between masking high gradient information to encourage diverse learning pathways, and masking low gradient information to reduce background noise and strengthen crucial synoptical pathways. The effectiveness of HLGM is validated through a unique metric that measures the alignment of high-fidelity saliency feature maps with labeled objects in images. Our comparative analysis against baseline models and current advanced techniques demonstrates substantial improvements in both model accuracy and saliency mapping. HLGM not only outperforms conventional training methods in accuracy but also advances model interpretability, positioning it as a pivotal tool in the pursuit of explainable AI in machine learning.

Submission Number: 182