Anticausal Learning for Inverse Problems and its Application on Optoacoustic Tomography

Artificial intelligence algorithms commonly exhibit poor performance when deployed on data whose distribution deviates from the one utilized during the training phase. While this vulnerability can be addressed post-training, doing so may necessitate a computationally intensive fine-tuning process an...

Descripción completa

Guardado en:
Detalles Bibliográficos
Autores principales: Vera, Matías, González, Martín Germán, Rey Vega, Leonardo
Formato: Artículo publishedVersion
Lenguaje:Español
Publicado: FIUBA 2025
Materias:
Acceso en línea:https://elektron.fi.uba.ar/elektron/article/view/222
https://repositoriouba.sisbi.uba.ar/gsdl/cgi-bin/library.cgi?a=d&c=elektron&d=222_oai
Aporte de:
Descripción
Sumario:Artificial intelligence algorithms commonly exhibit poor performance when deployed on data whose distribution deviates from the one utilized during the training phase. While this vulnerability can be addressed post-training, doing so may necessitate a computationally intensive fine-tuning process and/or require a significant acquisition of new data. In this context, causality theory presents an excellent paradigm for distinguishing variation-prone mechanisms from invariant ones. This distinction would permit fitting the model exclusively to the variable components, thereby reducing the complexity of the overall problem. However, this paradigm remains under-explored in relation to inverse problems, primarily because such problems are, by their very definition, anticausal. This work undertakes an analysis of the performance and inherent limitations of fundamental algorithms in inverse problems that satisfy the criteria for anticausal learning. Specifically, these algorithms are investigated within the context of image reconstruction in optoacoustic tomography.