Dile: a distribution-based incremental learning approach

Eduardo F. Morales

doi:10.1007/s10044-025-01552-4

Return

Dile: a distribution-based incremental learning approach

2025-09-26

PRE

Eduardo F. Morales *

Javier Herrera-Vega

Jonathan Serrano-Cuevas

Yareli Aburto

Luis Enrique Sucar

Hugo Jair Escalante

DOI：10.1007/s10044-025-01552-4

Original

Original request for help

Save

Abstract

En 中文

Image classification has thrived with the arrival of Deep Learning. However, the learned models are mostly static (cannot handle new classes), require a large number of images and long training times; and, in general, heavily depend on the user who defines the classes and provides labeled training data. Recently, class incremental learning has gained popularity by removing the need to retrain a model once data from new classes is available. Most incremental systems know when new classes are given and focus on not decreasing their global performance as these become available. In this paper, we describe DILE, an incremental learner based on distributions of embedded features and their comparisons using the Fréchet distance. DILE uses a limited amount of memory per class by using statistics of these feature vectors. We also introduce an incremental tree-based representation that is gradually built as more classes are known to the system, and which significantly reduces the classification times while placing semantically closer classes in the same branches of the tree. We tested the proposed approach on several databases with very competitive performance against state-of-the-art systems. Additional experiments were performed to see how DILE can work with a single image per class and automatically obtaining additional data from the Internet, and how well DILE can work if it has no knowledge if the examples are from a known or unknown class.

Keywords:

Incremental learning

Distribution-based classifier

Fréchet distance

Continual learning