Tech & AI LIVE: Gen AI – Epistemic AI, Prof. Fabio Cuzzolin

December 05, 2024

undefined mins

Professor Fabio Cuzzolin from Oxford Brookes University presents his groundbreaking work on epistemic AI at Tech & AI LIVE: Gen AI

Professor Fabio Cuzzolin from Oxford Brookes University captivated the audience at Tech & AI LIVE: Gen AI, presenting his groundbreaking work on epistemic AI. With an emphasis on addressing uncertainty in machine learning, his innovative approach promises to redefine AI’s applications across industries.

What is epistemic AI?

Epistemic AI challenges traditional machine learning paradigms by introducing a new methodology: epistemic learning.

This approach, as Fabio explains, seeks to model the uncertainty inherent in data and predictions.

“Current AI struggles with uncertainty due to limited data, which restricts its generalisation power,” Fabio notes.

Traditional machine learning typically generates single predictive models. Epistemic AI, in contrast, uses training data to derive a set of potential models, acknowledging gaps in understanding and adapting more effectively to new or changing conditions.

This work is the result of a Horizon 2020 collaboration with institutions such as KU Leuven in Belgium and TU Delft in the Netherlands. The €3 million, five-year project employs advanced mathematical techniques like credal sets and random sets to represent and manage uncertainty.

Professor Fabio Cuzzolin from Oxford Brookes University

Practical innovations and breakthroughs

Fabio showcases several applications of epistemic AI developed by his team, including:

1. Random set neural networks

In a departure from standard neural networks, random set neural networks predict not just a class but a range of possible outcomes, allowing for more nuanced decision-making.

“These models are not only more accurate but also better at identifying outlier data and withstanding adversarial attacks,” Fabio explains.

2. Enhanced generative AI

Addressing the lack of diversity in outputs from models like Stable Diffusion, Fabio’s team has introduced variability control techniques. By modelling Gaussian noise distributions, their methods enable the generation of more diverse and creative content.

3. Epistemic large language models

Large language models are another focus of this work. Traditional models often suffer from ‘hallucinations’, generating false information. Fabio’s epistemic approach reduces such errors by predicting a range of plausible outcomes rather than a single answer, a significant step forward in improving AI reliability.

Professor Fabio Cuzzolin from Oxford Brookes University

Expanding horizons

The applications of epistemic AI extend far beyond traditional AI systems. Fabio outlines several domains ripe for disruption, including:

Autonomous vehicles: Developing robust perception systems that adapt to changing environments.
Surgical robotics: Enhancing precision and reliability in medical procedures.
Climate change prediction: Modelling complex systems with uncertainty for better forecasts.
AI for science: Tackling challenges like nuclear fusion control and causal explanations in physical systems.

The project is also paving the way for startups and partnerships. Fabio reveals plans to incubate a startup focused on applying epistemic AI to autonomous vehicle systems through a European Innovation Council transition grant.

The future of epistemic AI

When asked about the future, Fabio acknowledges that scaling epistemic AI depends on resources and support.

“We’re competing with giants like OpenAI and Microsoft, but our proof-of-concept results are compelling,” he says.

Scaling up will enable the technology to address critical real-world problems, from healthcare to environmental challenges.

Fabio concludes with a call to action: “Epistemic AI represents a more robust, adaptable, and reliable future for artificial intelligence. The potential applications are vast, and we’re just beginning to scratch the surface.”