Juan Sebastian Olier
Wednesday November 14, 14.30h, Top Floor, Deprez
Perceiving the world implies interpreting or making sense of the content of sensory inputs. Defining a mechanism to achieve such a process is a central task to create learning algorithms and embed perceptual capabilities into artificial agents. Prominent ideas in the cognitive sciences emphasize the dynamic nature of perception and the need for it to stay grounded in the sensorimotor system. Thus, a mechanism for interpreting the content of sensory stimuli has to comply with such dynamism and align with views of grounded and embodied cognition. Those requirements are sometimes viewed as irreconcilable with the idea of representations.
Nonetheless, the postulates of Active Inference and Predictive Processing can be argued to be in line with the mentioned requirements, but also to constitute a framework to reconcile them with the notion of representations. In particular, I claim that such framework enables a dynamic interpretation of representations in the perceptual process, and thus to overcome the limitations that a view of representations as static structures might bring about. With the aim of translating such approaches to implementable algorithms, I have made a connection between the ideas behind Active Inference and the state of the art on Generative Models in Deep Learning. From that analysis, I have developed learning models for building dynamic representations.
In this talk, I will elaborate on these arguments, and will present some of the Deep Learning models that I have proposed, along with some results; Especially, I will develop the connection between Deep Generative Models and the idea of Dynamic Representations.