Resumen de Ikaskuntza autogainbegiratuaren ahalmena azaleratzen

Aitor Sánchez Ferrero, Borja Calvo Molinos , Usue Mori Carrascal , José Antonio Lozano Alonso

• euskara

  Ikaskuntza automatikoak nabarmen egin du aurrera azken urteotan, eta horren fruitu da garatutako algoritmo sorta izugarria, zeinek ahalmena ematen baitigute ataza mota ugari egiteko. Literaturaren arabera, algoritmo gehienak ikaskuntza gainbegiratuan oinarritzen dira. Hala ere, hainbat atazatan emaitza onak lortu arren, ikaskuntza-paradigma horren eragozpen nagusia etiketetiko mendekotasuna da; izan ere, etiketatze-prozesua oso garestia da. Gainera, ikaskuntza automatikoan erabiltzen diren ereduek joera handia dute alborapen okerrak bultzatzen dituzten bide laburrak ikasteko, eta, ondorioz, atazek porrot egiten dute. Arazo horiek saihesteko, ikaskuntza autogainbegiratuak arreta bereganatu du azkenaldian ikaskuntza-paradigma gisa. Lan honetan, ikaskuntza autogainbegiratuari dagokion literaturari buruzko sarrera bat eskaintzen dugu, eta ikaskuntza-paradigma horren barruan bereizten diren metodo motak biltzen eta azaltzen ditugu, zenbait datu motatan aplikatzeko oinarrizko prozedurak aztertuz.

• English

  Machine learning has made significant progress in recent years, resulting in the development of an impressive array of algorithms that enable us to perform a wide variety of tasks. According to the literature, most algorithms rely on supervised learning. However, despite achieving good results in different tasks, the main drawback of this learning paradigm is its dependency on manually created human labels, as the labeling process is very costly. Moreover, machine learning models tend to learn shortcuts that promote incorrect biases, leading to failures in the tasks they aim to accomplish. To avoid these issues, self-supervised learning has recently gained attention as a learning paradigm. This work provides an introduction to the literature on self-supervised learning. It explains the different types of methods distinguished within this learning paradigm and examine the basic procedures for applying them to different types of data.