Continual Learning applications

Introduction

Continual Learning (CL) aims at defining models that continuously learn and evolve according to new amounts of data, retaining previously learned knowledge.

• The model is able to incrementally learn and autonomously change its behaviour without forgetting the original task.

• Mitigate catastrophic forgetting of old concepts when new ones are learned.

• Learn a new task without using training data for previous tasks, which might no longer be accessible.

Our Research

We started by studying CL for classification of biomedical images: Colon Pathology, Dermatoscope, Retinal OCT, Blood Cell Microscope and Kidney Cortex Microscope. In particular, we compared different CL strategies (i.e.,  Naive, Replay, CWR*, ICaRL, and Cumulative approaches) to support DL architectures in classifying medical images.

Publications

[1] A. Quarta, P. Bruno, and F. Calimeri, "Continual Learning for medical image classification", in CEUR Workshop Proceedings, vol. 3307, pp. 67-76, 2022. 1st AIxIA Workshop on Artificial Intelligence for Healthcare, HC@AIxIA 2022, 30 November 2022.