Título: HebbAL: A Hebbian -Based Strategy for Kernel Active Learning
Autores: Yan V. G. Ferreira, Elias José de Rezende Freitas, Luiz Carlos Bambirra Torres & Antonio de Pádua Braga
Resumo: Although most machine learning methods rely on the availability of large sets of labeled training instances, most real-world available data is unlabeled, since labeling can be expensive and time-consuming. To avoid the expense of labeling all the data before fitting a model, the paradigm of active learning allows models to query the labeling of only the most relevant samples to its training. However, when active learning is performed online, specially with streaming data, updating the model can be challenging, as it usually requires running several optimization steps or inverting matrices to retrain a model. In fact, this can make learning unfeasible for systems with computational constraints, such as embedded hardware. In this context, the present work aims at enhancing active learning with Hebbian classifiers, which allow updating param- eters inexpensively to learn new samples or to remove outdated information. For this purpose, we employ a recently proposed kernel-based method for embedding data in a suitable way for Hebbian learning and combine it with a querying criterion based on the Perceptron Convergence Theorem. The resulting model, called HebbAL, is applied to 17 UCI benchmarks, presenting data to the model in a streaming fashion. Our approach achieved comparable performance to widely used active learning models, such as a Random Forest with uncertainty querying. Although HebbAL presents a slightly lower AUC on average, it queries less labels than the other models. Therefore, our novel approach to online active learning represents a step towards building scalable learning systems to act in dynamical environments.
Palavras-chave: Active learning; Hebbian learning; kernel learning; online learning.
Páginas: 8
Código DOI: 10.21528/CBIC2025-1175647
Artigo em PDF: CBIC_2025_paper1175647.pdf
Arquivo BibTeX:
CBIC_2025_1175647.bib