Not Knowing in Deep Representation Learning

Descripción del proyecto

Resolver el reto de la «caja negra»

La mejora del aprendizaje automático y la inteligencia artificial lleva aparejado un aumento de la complejidad de los modelos, lo que los hace cada vez más opacos a la comprensión humana. Este fenómeno se conoce como el problema de la «caja negra». La opacidad surge de un problema de identificabilidad, en el que los modelos pueden expresar el mismo patrón de múltiples maneras diferentes dentro de sus representaciones internas. La falta de transparencia obstaculiza nuestra capacidad para comprender la información generada por los modelos, lo cual plantea un dilema: aceptar la inescrutabilidad de los resultados o sacrificar la complejidad del modelo. En este contexto, el proyecto NoKnow, financiado por el Consejo Europeo de Investigación, desplazará el foco de atención desde las representaciones a las tareas que estas resuelven. El equipo del proyecto NoKnow empleará herramientas de geometría diferencial e inferencia bayesiana para garantizar resultados identificables dentro de la representación múltiple.

Objetivo

Machine learning and artificial intelligence techniques are progressing at a tremendous pace and impressive results appear across scientific fields. However, as machine learning models grow in capacity, they become increasingly ‘black box’, and it becomes harder for humans to reason about the patterns discovered by the machine. A root cause of this difficulty is that most machine learning models can express the same pattern in infinitely many, equally good, ways within their internal representations of the world. This is known as an identifiability problem.

Today we lack a general solution to identifiability problems, and either give up on understanding the patterns discovered by the machine or reduce model complexity to lessen the problem. The latter also reduces the fidelity and applicability of the model. NoKnow rephrase the question of identifiability to be concerned with tasks solved by the representation rather than the representation itself. Using tools from differential geometry and Bayesian inference, we develop the theoretical tools to ensure that tasks solved in the representation have an identifiable outcome even if the representations themselves are not identifiable.

To turn theory into practice, we develop state-of-the-art algorithms for assessing the uncertainty of learned representations in order to indirectly estimate the topology of the representation manifold. We further develop novel, high-fidelity predictive models that have identifiable outcomes when trained on learned representations.

NoKnow provides the fundamental tools needed to engage with learned representations to guarantee identifiable outcomes. This, in turn, increase trust in findings as they are not dependent on arbitrariness in learned representations. As society increasingly automates decisions, this trust in machine learning becomes ever more important.