Curriculum learning: scheduling training data

In a typical educational system, learning relies on a curriculum that introduces new concepts building upon previously acquired ones. The rationale behind, is that humans and animals learn better when information is presented in a meaningful way rather than randomly. We follow this starting small concept to design our approaches with convolutional neural networks for medical image classification tasks.

I developed a novel curriculum learning (CL) strategy to improve femur fracture classification from X-rays [JA19], showing that training sample order significantly impacts model performance, especially in medical contexts where domain knowledge is key. By integrating domain knowledge from medical guidelines, expert annotations, and ambiguities in their annotations, I improved model reliability and classification accuracy. I further extended this work into a comprehensive CL framework with three strategies: sample weighting, training set ordering, and subset sampling [JA22]. Additionally, I showed that model uncertainty can serve as a fallback when domain knowledge is unavailable.

To culminate this line of research, I designed a federated learning (FL) framework that integrates CL and unsupervised domain adaptation to improve breast cancer classification across multi-centers datasets [JA23a]. FL offers a secure and privacy preserving method to train a ML algorithm across multiple decentralized nodes that keep their data samples locally. In this setting, local models perform computation on their private data to update the global model. Our CL approach utilized a data scheduler to prioritize local training samples, with the objective of reinforcing local model consistency by penalizing forgotten samples.

Publications

[JA23a] Memory-aware curriculum federated learning for breast cancer classification
Amelia Jiménez-Sánchez, Mickael Tardy, Miguel A. González Ballester, Diana Mateus, Gemma Piella
Computer Methods and Programs in Biomedicine 2023
PDF   Bibtex  

[JA22] Curriculum learning for improved femur fracture classification: scheduling data with prior knowledge and uncertainty
Amelia Jiménez-Sánchez, Diana Mateus, Sonja Kirchhoff, Chlodwig Kirchhoff, Peter Biberthaler, Nassir Navab, Miguel A. González Ballester, Gemma Piella
Medical Image Analysis 2022
PDF   Bibtex   Code

[JA19] Medical-based deep curriculum learning for improved fracture classification
Amelia Jiménez-Sánchez, Diana Mateus, Sonja Kirchhoff, Chlodwig Kirchhoff, Peter Biberthaler, Nassir Navab, Miguel A. González Ballester, Gemma Piella
MICCAI 2019
PDF   Bibtex   Slides   Poster

Funding

• European Union’s Horizon 2020 research and innovation programme under the Marie Skłodowska-Curie grant agreement No. 713673.
• “la Caixa” Foundation (ID Q5850017D), fellowship code: LCF/BQ/IN17/11620013.
• Spanish Ministry of Economy [MDM-2015-0502].