Deep Inference

This presentation considers deep models in the brain. It builds on previous formulations of active inference to simulate behaviour and electrophysiological responses under deep (hierarchical) generative models of discrete state transitions. The structured temporal aspect of these models means that evidence is accumulated over time, enabling inferences about narratives and policies. We illustrate this behaviour in terms of Bayesian belief updating – and associated neuronal processes – to reproduce epistemic foraging for reward. These simulations reproduce these sort of delay period activity and local field potentials seen empirically; including evidence accumulation, place cell activity and transfer of dopamine responses. These simulations are presented as an example of how to use basic (first) principles to constrain our understanding of functional architectures that underwrite natural and artificial intelligence.

People Before Algorithms: Building a Trustworthy AI for a Better Future

GenAI powered Contextual Search for E-commerce

Leveraging foundation models for the discovery of novel trustworthy algorithms

Innovating with AI: The Next Wave of Entrepreneurship

Navigating the AI Technology Stack: A Global Overview Through a Japanese Lens

Assessing Morocco's AI Readiness

In May 2024, UNESCO published the most comprehensive report on Artificial Intelligence in Morocco, through 5 dimensions: legal and regulatory, social and cultural, economic, scientific and educational, and technological and infrastructural. With 17 recommendations, UNESCO provided a Roadmap to successfully advance AI in the country, through the lens to the Recommendation on the Ethics of AI

Advancing the Fourth Paradigm: Machine Learning Frameworks for Experimental Science

Scientific breakthroughs in the fourth paradigm of discovery are increasingly driven by advanced computing capabilities that enable researchers to manipulate, explore, and exploit large datasets. Machine learning has accelerated the pace of scientific discovery across various domains, offering new tools and methods throughout the main steps of the experimental science loop. However, current ML approaches face many limitations in scientific applications, particularly in drug discovery and materials science, where the exploration of vast chemical and material spaces remains challenging. To address these shortcomings, we examine generative flow networks (GFlowNets), a novel machine learning framework for systematic scientific exploration and discovery.

AI, Wittgenstein and the wonders of Poetics

The essence of the world is logico-mathematical: AI can perfectly penetrate that. The non-essential part of the world (poetics) is only meaningful at the individual level and it supposes communication between two individuals. It can be mimicked by AI but it misses the core of the message.

AI for Disaster Response.

Natural disasters often leave widespread devastation in their wake, making timely damage assessment critical for effective response and recovery efforts. This presentation explores SKAI, a Google-developed framework that leverages AI and machine learning to accelerate and enhance the damage assessment process. By combining computer vision models and large vision language models, SKAI aims to improve both the speed and coverage of post-disaster evaluations. This presentation will delve into the technical aspects of SKAI, illustrating how it can be a valuable tool for decision-makers in the crucial aftermath of a disaster. The potential implications of such a tool for disaster response and recovery will be discussed, aiming to engage a mixed audience of technical experts and those interested in the broader applications of AI for social good.

Innovation in the Age of AI

- Why innovation is critical for compagnies today, in the age of digital and AI
- ⁠Future AI trends and how they will shape our future
- ⁠How compagnies can leverage AI as an innovation lever, at the core of their growth strategy
- ⁠Business & country use case : AI as an accelerator of Morocco’s development for a successful WorldCup 2030 & beyond

Governing with AI

Countries are increasingly exploring the potential of Artificial Intelligence (AI) to enhance public governance. AI applications in the public sector offer opportunities to boost productivity, improve the responsiveness of public services, and reinforce government accountability. However, these advancements also present significant risks, calling for the creation of an enabling environment for trustworthy AI. This presentation examines the key trends and policy challenges related to the development, deployment, and use of AI in the public sector. It highlights the benefits and risks of AI adoption, showcases its role in enhancing efficiency and accountability, and provides insights into how OECD countries are addressing critical policy issues through innovative practices and strategies.

AI-Aided Advisors: Design and Validation

We have developed a site called dietnerd.org which answers questions about diet and health. The workflow starts with a query, then proceeds through parsing, pointing at a database, relevance analysis, safety analysis, and summarization. The challenge is to ensure that the result is clear and accurate. This talk presents our design choices (to privilege safety and to avoid hallucinations) and our validation tests (blinded studies with human experts as well as comparison with systematic reviews). I will discuss various generalizations of this approach to AI-based experts on the wireless literature and standards and to financial investment.

Tabular learning, from data preparation to foundation models

While much of the excitement is on AI generating images, much of the data-science challenges are about assembling and preparing tabular data. These typically require extensive manual transformation or "data wrangling", and gradient-boosted trees are king. I will discuss how we progressively rethought this process, eventually laying the ground of foundation models for tabular data.