Efficient Single-Step Framework for Incremental Class Learning in Neural Networks

Alejandro Dopico-Castro · 2025

Read Efficient Single-Step Framework for Incremental Class Learning in Neural Networks by Alejandro Dopico-Castro (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-2509-11285.txt.

Description

Teaching a neural network new classes over time without erasing old ones is the problem behind this 2025 machine learning paper by Alejandro Dopico-Castro and three co-authors. The usual approach, repeated gradient training, is costly, touchy about settings, and prone to catastrophic forgetting.

Their framework, CIFNet, takes a different route: with a frozen pre-trained feature extractor, each new class can be added by an exact closed-form ridge regression update, with no backpropagation at all. A small calibration buffer in feature space, plus density-aware oversampling, keeps the decision borders fair between old and new classes. On three standard benchmarks the method matched iterative rivals while learning smoothly, never collapsing midway, and using far less energy. The results make a case for analytic, formula-based continual learning.

Catalog record

CREATORAlejandro Dopico-Castro
YEAR2025
COLLECTIONPAPERS
FORMATTXT
FILE SIZE53 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2509-11285

Browse the full library · About the Tower of Babel