AI systems are often a mirror of human behavior, and what we see in this mirror is not always pleasant. From our behavior and the data we produce, AI systems learn to make decisions just like a human - and so these decisions sometimes turn out to be just as discriminating and unfair as those made by humans. Ensuring fairness in AI decision-making is arguably one of the most significant current challenges in computer science. Major concerns are also how to explain the machine's decision-making process and how to protect personal data used to train the AI.
But how can these daunting requirements be translated into verifiable properties of computer programs? "Explainability or fairness are not well-defined mathematical properties, but complex social concepts. Our goal is not to define once and for all what fairness is, for example. This question has hugely complex philosophical and societal dimensions. We want to allow researchers and developers to express precisely what they mean for each individual program to be fair. Just saying 'it should be fair' leaves 1000 plausible interpretations," explains Bernd Finkbeiner.
This is why the CISPA researcher is working on a specification language in the HYPER project that lets developers describe exactly how they intend their software to behave. Specification languages allow users to describe what a program should do at a higher level. Thus, a logical and consistent model can be developed, which is later translated into program code.
To be able to express and describe properties such as explainability, fairness, and privacy in such a way that they can be embedded in different programs and also proven with different analysis techniques, a unifying theory has been lacking up to now, according to Bernd Finkbeiner. The reason for this is the complexity of these properties, which are called hyperproperties. "They are much more expressive than other properties that are traditionally used to characterize the correctness and reliability of computer programs. Hyperproperties can express relations between different situations." For example: In a program, the principle 'men and women are treated equally' should apply. Compliance with this principle cannot be tested on isolated program executions but only by analyzing all program executions and comparing them with each other. In this context, a program execution is the sequence of outputs a program produces in response to the inputs of a male or female user.
"Previous attempts to capture hyperproperties in a specification language for software systems have only been successful for a narrow section of the entire spectrum of hyperproperties," says Bernd Finkbeiner. His project aims to change that. Once a common specification language has been found, the next step in the project will be to develop new algorithms. They will then monitor, verify, and synthesize programs. "Monitoring means we analyze the data produced by a program and check whether the desired (hyper-) properties actually hold or are violated. Verification refers to finding problems in the program code. Another important goal is synthesis, the automated construction of a program that can be shown to have the desired properties."
Ultimately, Finkbeiner says, this also enables measuring the project's success by having researchers incorporate the methods into real-world applications. "Essentially, my research is then used wherever humans hand over decision-making responsibility to machines. This includes all autonomous areas such as self-driving vehicles, unmanned drones, or application management systems that help decide who gets a job,” explains the senior scientist.
Bernd Finkbeiner has previously already been awarded an ERC Consolidator Grant. The Advanced Grant is given to experienced researchers with significant research results over the past ten years. For the researcher, this is a great honor. "The Advanced Grant gives me as a researcher the opportunity to implement pioneering ideas - even with the risk of failure. The funding will mainly go towards hiring young researchers, who will support me in implementing the project."
About the ERC
The ERC, set up by the European Union in 2007, is the premier European funding organisation for excellent frontier research. It funds creative researchers of any nationality and age, to run projects based across Europe. The ERC offers 4 main grant schemes: Starting Grants, Consolidator Grants, Advanced Grants and Synergy Grants. The ERC is led by an independent governing body, the Scientific Council.