Orthogonal Activation with Implicit Group-Aware Bias Learning for Class Imbalance

1 st Sukumar Kishanthan, 2 nd Asela Hevapathige · 2025

Read Orthogonal Activation with Implicit Group-Aware Bias Learning for Class Imbalance by 1 st Sukumar Kishanthan, 2 nd Asela Hevapathige (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2512-20006.txt.

Description

Sukumar Kishanthan and Asela Hevapathige redesign the activation function to fight class imbalance. When training data favor one class by far, classifiers learn to neglect the rare class. Most fixes resample the data first, or adjust the outputs last. Their activation, OGAB, steps in during training. It acts where the embeddings form. It needs no explicit label facts to do so.

Two tools do the work. Orthogonal maps keep features independent. That saves rare-class facts that majority rule would crush in the embedding space. A group-aware bias tool sniffs out natural clusters in the data. It then shifts embeddings to widen the gaps between classes on its own. On real and synthetic imbalanced data sets, OGAB beat classic and learnable rivals. The gains were steady across the tests.

Catalog record

CREATOR1 st Sukumar Kishanthan, 2 nd Asela Hevapathige
YEAR2025
COLLECTIONPAPERS
FORMATTXT
FILE SIZE22 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2512-20006

Browse the full library · About the Tower of Babel