• Baptiste Caramiaux
  CNRS Researcher ISIR, Sorbonne Université 

Interactivity in Machine Learning
Machine learning (ML) has shown remarkable results over the last ten years and has been integrated in many user-facing applications. The vast majority of ML methods are, however, fixed systems when deployed: users have no way of explicitily modifying the behaviour of the underlying models that make recommendations or predictions. In this context, research at the intersection of human-computer interaction and ML has proposed 'Interactive Machine Learning' (IML) as the approach where users, facing ML models, have access to the data and the learning model, so that they can act on it to make it more robust for specific application cases, more personalised and more trustworthy. In this presentation, I propose to come back to the promises of this approach, which aimed to give power back to people, to present where we are today in the field and where we’re going (or can go). I will illustrate the presentation with work from my research group in the fields of creative and artistic applications.

• Rufin Vanrullen 
  Research Director, Artificial & Natural Intelligence Toulouse Institute CNRS, CerCo

The Global Latent Workspace: a model of cognition with AI applications
Global Workspace Theory (GWT) is a leading account of human cognition and consciousness. In this view, a number of independent specialized modules connect to a shared central representation space; when a module is selected by attention, its contents are mobilized into the Global Workspace, and broadcast across the entire brain, resulting in a unified and integrated experience. Inspired by this framework, we have developed a deep learning architecture that captures key features of GWT: the Global Latent Workspace (GLW). I will present our GLW and its initial implementations, with promising applications in various AI domains. The model shows improvements in sample efficiency for multimodal representation learning. It can be leveraged for downstream classification and retrieval tasks. When an action module is connected to the GLW, the whole system exhibits affordance-like properties. The GLW is also beneficial as an input space for RL policy training: the policy is learned with fewer environment steps, and displays zero-shot cross-modal transfer abilities. Finally, augmenting the GLW with “operation” modules and an attention-controlled routing mechanism could open the way toward System-2 reasoning and sequential problem-solving. As GWT is also widely held as a neuroscientific theory of consciousness, I will finish by discussing the possible implications of these systems for AI consciousness.

Rufin VanRullen is a CNRS Research Director at CerCo (Toulouse) and holds an AI Research Chair at ANITI (Artificial and Natural Intelligence Toulouse Institute). He leads the ERC Advanced GLoW project, exploring novel deep-learning architectures based on the Global Workspace Theory. He has published more than 150 scientific papers, including influential articles on neural coding, object recognition, feed-forward vs. feedback processes, and attention.

• Raja Chatila
  Professor, ISIR, Sorbonne Université

Self-Awareness: A constructivist Approach
(Self-)consciousness, sometimes equated with self-awareness, is an ill-defined concept. It has various definitions in Philosophy, in Psychology and in Neurosciences. There are several intuitive concepts to designate it, and theories to model it. We aim to investigate if we can frame it from a computational perspective.
We consider therefore that a conscious entity must be embodied, capable of perception and action in the environment so that it can have its own experience of the world. It must also have the capability of interpreting its interactions with respect to itself, distinguishing the effects of its actions from changes in the environment produced by other causes. But this is not sufficient. Built on these features, we also hypothesize that it must be able to self-reflect and to assess its knowledge to make decisions, and to augmented its capabilities.
To find out if these hypotheses are workable to achieve consciousness, our approach is to construct and experiment computational mechanisms and systems integrated in a global architecture and embodied in a robot able to interact with its environment and other agents, exhibiting the mentioned capabilities.