همایش بین المللی هوش مصنوعی

صفحه اصلی / دومین همایش بین المللی هوش مصنوعی

SB Softmax: A Parametric and Robust Replacement for Softmax in Deep Learning Models

نویسندگان :

Morteza Taheri¹

1- دانشگاه زنجان

کلمات کلیدی :

SB-Softmax،PSO،calibration

چکیده :

Abstract—The conventional Softmax activation often exhibits numerical instability and overconfident predictions arising from its unbounded exponential kernel, particularly under mixed precision training. This paper presents SB Softmax (Sigmoid Bound Softmax), a bounded and parametric normalization layer defined by learnable parameters α and β, initialized via Particle Swarm Optimization (PSO). The proposed formulation preserves the normalization structure of Softmax while providing self bounded derivatives, ensuring stable gradient flow and maintaining O(N) computational complexity. Experimental evaluations conducted on CIFAR 10, MNIST, and CIFAR 100 demonstrate consistent improvement in accuracy and calibration metrics with reduction in training time, compared to the standard Softmax. Specifically, accuracy improved from 0.7736 to 0.7882 on CIFAR 10 and from 0.4576 to 0.4887 on CIFAR 100. These results confirm the effectiveness of SB Softmax as a robust and efficient replacement for Softmax in deep learning models.

لیست مقالات

لیست مقالات بایگانی شده

Fibroglandular Tissue Classification in Breast MRI: A Comparative Study of Automated Decision Strategies

Meysam Khalaj - Arvin Arian - Ala Torabi - Nasrin Ahmadinejad - Masoumeh Gity - Seyedeh Nooshin Miratashi Yazdi - Mohammad Pooya Afshari - Melika Sadeghi Tabrizi - Hamid Soltanian-Zadeh

An interpretable framework based on deep learning and the Internet of Things for predicting risk dynamics in chronic patients

Behnaz Pouriayevali - َAsghar Ehteshami

Interpretable Machine Learning for Rocking-Induced Settlement Prediction Using SHAP Analysis

Seyed Emad Miri - Hamid Mohammadnezhad