Periodic Reporting for period 2 - E-pi (Epistemic AI)
Okres sprawozdawczy: 2022-03-01 do 2024-08-31
Although artificial intelligence (AI) has improved remarkably over the last years, its inability to deal with fundamental uncertainty severely limits its application. Current AI models cannot confidently make predictions robust enough to stand the test of data which differs (often even by tiny details) from the data used to train them. As a result, AI systems suffer from brittle behaviour, and find it difficult to operate in new situations.
Epistemic AI’s objective is to give birth to a new paradigm for a next-generation artificial intelligence able to provide robust, worst-case guarantees on its predictions, thanks to a proper mathematical modelling of real-world ‘epistemic’ uncertainty, that is, uncertainty about the very nature of the process generating the data.
This new learning paradigm can pave the way for intrinsically safe autonomous vehicles, next-generation human-robot interaction systems, more efficient manufacturing techniques, increased fairness and trustworthiness of AI algorithms.
The project aims to formulate a new mathematical framework for optimisation under epistemic uncertainty, laying, in turn, the premises for the creation of new ‘epistemic’ learning paradigms. Epistemic AI focusses, in particular, on some of the most important areas of machine learning: unsupervised learning, supervised learning and reinforcement learning.
The project’s ambition is also to foster an ecosystem of academic, research, industry and societal partners throughout Europe able to drive and sustain the EU’s leadership aspirations in the search for a next-generation AI.
Epistemic AI’s objective is to give birth to a new paradigm for a next-generation artificial intelligence able to provide robust, worst-case guarantees on its predictions, thanks to a proper mathematical modelling of real-world ‘epistemic’ uncertainty, that is, uncertainty about the very nature of the process generating the data.
This new learning paradigm can pave the way for intrinsically safe autonomous vehicles, next-generation human-robot interaction systems, more efficient manufacturing techniques, increased fairness and trustworthiness of AI algorithms.
The project aims to formulate a new mathematical framework for optimisation under epistemic uncertainty, laying, in turn, the premises for the creation of new ‘epistemic’ learning paradigms. Epistemic AI focusses, in particular, on some of the most important areas of machine learning: unsupervised learning, supervised learning and reinforcement learning.
The project’s ambition is also to foster an ecosystem of academic, research, industry and societal partners throughout Europe able to drive and sustain the EU’s leadership aspirations in the search for a next-generation AI.
In this first two reporting periods, concerning epistemic learning theory, we reviewed the state-of-the-art in optimisation under uncertainty and devised optimisation frameworks for classification under epistemic uncertainty, developed a method to estimate interval uncertainty in surrogate models, worked to extend epistemic learning to neuro-symbolic approaches, devised an evaluation framework for uncertainty-aware predictions, and explored alternative approaches to learning a random set on a network’s parameter space.
In the statistical domain, we laid the groundwork for a new epistemic learning theory, assuming that train and test data arise from a convex set of distributions, deriving realistic generalisation bounds effective under epistemic uncertainty.
Within unsupervised learning and generative AI, we proposed a new deep evidential clustering approach, a novel “epistemic diffusion model” concept able to generate more diverse data, and applied the concept of random-set prediction to large language models capable of more diverse sentence generation and less prone to hallucination.
In the supervised domain, we formulated of an original random-set neural network approach and a new credal interval neural network architecture, formulated credal deep ensembles, developed a framework for imprecise evidential classification and a “credal wrapper” concept mapping the outputs of Bayesian neural networks and deep ensembles to a credal set.
Within reinforcement learning (RL), we designed a novel distributional RL algorithm that jointly addresses aleatoric and epistemic uncertainty, studied the use of RL to solve scheduling problems and extended Monte-Carlo tree search to include epistemic uncertainty; we introduced a Sequential Monte-Carlo method for Bayesian Q-Learning and worked to set up a benchmark dataset to test inverse reinforcement learning abilities under uncertainty.
Significant progress has also been done in our validation domain (autonomous driving), by defining our science-to-technology use cases, creating and refining a Data Management Plan and a Fallback and General Safety plan. Research-wise, we released a ROad event Awareness Dataset for autonomous driving (ROAD) through a challenge at ICCV 2021, developed a framework for multi-objective epistemic reinforcement learning and experimentally tested our epistemic classifiers for weather condition and cone classification; we developed and tested uncertainty-aware object detectors, devised a graph network for modelling complex activities detection in video; we further extended the ROAD dataset to both the Waymo dataset and videos captured in the UAE, while organising challenges and competitions at ICCV’23, ECCV’24 and IROS’24 on this topic.
We have also engaged numerous partners to build a European ecosystem around epistemic AI, including the Universities of Pennsylvania, Manchester, Edinburgh, Eindhoven, Bristol, Cambridge and Imperial College London, Khalifa University in Abu Dhabi, National Yang Ming Chiao Tung University and IIT Bombay, a consortium of 10 UK universities on the creation of an AI Hub, companies such as Alien Technology Transfer, Perceive.ai Leo Drive, Createc, NVIDIA, Zebra Technologies, ORTEC as partners supporting ongoing grant applications, a consortium of commercial partners in a proposal for sustainability at the Port of Rotterdam, a group of Dutch universities in a proposal that builds on E-pi contributions to epistemic reinforcement learning, a consortium (Pisa, UvA Amsterdam, the Helmholtz Centre for Environmental Research) towards a new EiC PAthfinder project on time-varying machine learning under uncertainty, another consortium (TU Dortmund, Sorbonne, UKAEA) on new uncertainty-aware neural operators for nuclear fusion plasma control, two more large consortia on bids on AI robustness and explicability and multimodal foundation models.
An Industrial Advisory Board was set to advise on the project’s exploitation plan, and possibly Key Exploitable Results have been identified.
In the statistical domain, we laid the groundwork for a new epistemic learning theory, assuming that train and test data arise from a convex set of distributions, deriving realistic generalisation bounds effective under epistemic uncertainty.
Within unsupervised learning and generative AI, we proposed a new deep evidential clustering approach, a novel “epistemic diffusion model” concept able to generate more diverse data, and applied the concept of random-set prediction to large language models capable of more diverse sentence generation and less prone to hallucination.
In the supervised domain, we formulated of an original random-set neural network approach and a new credal interval neural network architecture, formulated credal deep ensembles, developed a framework for imprecise evidential classification and a “credal wrapper” concept mapping the outputs of Bayesian neural networks and deep ensembles to a credal set.
Within reinforcement learning (RL), we designed a novel distributional RL algorithm that jointly addresses aleatoric and epistemic uncertainty, studied the use of RL to solve scheduling problems and extended Monte-Carlo tree search to include epistemic uncertainty; we introduced a Sequential Monte-Carlo method for Bayesian Q-Learning and worked to set up a benchmark dataset to test inverse reinforcement learning abilities under uncertainty.
Significant progress has also been done in our validation domain (autonomous driving), by defining our science-to-technology use cases, creating and refining a Data Management Plan and a Fallback and General Safety plan. Research-wise, we released a ROad event Awareness Dataset for autonomous driving (ROAD) through a challenge at ICCV 2021, developed a framework for multi-objective epistemic reinforcement learning and experimentally tested our epistemic classifiers for weather condition and cone classification; we developed and tested uncertainty-aware object detectors, devised a graph network for modelling complex activities detection in video; we further extended the ROAD dataset to both the Waymo dataset and videos captured in the UAE, while organising challenges and competitions at ICCV’23, ECCV’24 and IROS’24 on this topic.
We have also engaged numerous partners to build a European ecosystem around epistemic AI, including the Universities of Pennsylvania, Manchester, Edinburgh, Eindhoven, Bristol, Cambridge and Imperial College London, Khalifa University in Abu Dhabi, National Yang Ming Chiao Tung University and IIT Bombay, a consortium of 10 UK universities on the creation of an AI Hub, companies such as Alien Technology Transfer, Perceive.ai Leo Drive, Createc, NVIDIA, Zebra Technologies, ORTEC as partners supporting ongoing grant applications, a consortium of commercial partners in a proposal for sustainability at the Port of Rotterdam, a group of Dutch universities in a proposal that builds on E-pi contributions to epistemic reinforcement learning, a consortium (Pisa, UvA Amsterdam, the Helmholtz Centre for Environmental Research) towards a new EiC PAthfinder project on time-varying machine learning under uncertainty, another consortium (TU Dortmund, Sorbonne, UKAEA) on new uncertainty-aware neural operators for nuclear fusion plasma control, two more large consortia on bids on AI robustness and explicability and multimodal foundation models.
An Industrial Advisory Board was set to advise on the project’s exploitation plan, and possibly Key Exploitable Results have been identified.
Epistemic AI radically departs from current AI, as the latter fails to recognise the foundational issue the discipline has with the representation of uncertain knowledge. Our work also goes beyond ‘human-centric’ AI, since it strives to model the uncertainty stemming not just from human behaviour, but from all sources of uncertainty present in complex environments.
We expect the project to deliver a battery of theoretical frameworks and algorithms in all the main areas of machine learning, to be offered to practitioners in all fields of data science and applied science together with a wealth of publicly available codebases and datasets, proof-of-concept validation in autonomous driving scenarios, and a significant ecosystem of universities and companies sharing the ambition to unlock a new generation of uncertainty-aware AI.
Epistemic AI’s impact, while already appreciable in the medium term, will fully unfold in the longer term. The new epistemic learning paradigm can lead, within autonomous driving, to much improved safety; within robotics, it may enable robots with more robust perception and decision-making abilities. In manufacturing, E-pi’s algorithms may lead to more robust scheduling, resulting in increased production plant performance. An AI with an increased awareness of human behaviour will also be crucial to bring about fully autonomous robotic surgery.
We expect the project to deliver a battery of theoretical frameworks and algorithms in all the main areas of machine learning, to be offered to practitioners in all fields of data science and applied science together with a wealth of publicly available codebases and datasets, proof-of-concept validation in autonomous driving scenarios, and a significant ecosystem of universities and companies sharing the ambition to unlock a new generation of uncertainty-aware AI.
Epistemic AI’s impact, while already appreciable in the medium term, will fully unfold in the longer term. The new epistemic learning paradigm can lead, within autonomous driving, to much improved safety; within robotics, it may enable robots with more robust perception and decision-making abilities. In manufacturing, E-pi’s algorithms may lead to more robust scheduling, resulting in increased production plant performance. An AI with an increased awareness of human behaviour will also be crucial to bring about fully autonomous robotic surgery.