Name of the participant: Edoardo Mosca
Description of the IT research project: AI-driven NLP systems are powerful decision-makers and are at the core of the current technological revolution in science, business, and even in most aspects of our private lives. However, the machine learning models employed are opaque, which prevents their deployment despite their high performance. This is particularly true for high-stake applications, where disastrous consequences could originate if these systems do not operate as desired. At the same time, while automated agents outperform humans in some specific tasks, they still lack most of the generalization and reasoning capabilities that humans possess.
Working towards Strong AI and more accurate NLP models has been the main focus in academia and industry. This project abandons this core ideal and instead works towards combining the available technologies with human support to build stronger decision-makers. When it comes to deploying AI systems into the real world, we believe that enabling human oversight is a much stronger asset than squeezing an extra 1% performance out of the already complex models employed. This paradigm change can be a driving force for the larger adoption and acceptance of these new technologies in real-world applications, industry, and society.
Our final goal is to build a modular interface that enables a continuous and real-time interaction between an NLP agent and human stakeholders. While the agent can be any machine learning model suitable for the task at hand, the interaction with the stakeholders is driven by the latest Explainable AI techniques. Explanations provided by the agent are editable by humans in a user-friendly manner and can be fed back into the system to adjust its reasoning and influence its outcome. Multiple stakeholders can participate in the process as (1) their differences in expertise are considered and (2) their influence on the NLP agent is held accountable, making automatic decision-making a transparent and democratic process.
Software Campus partners: TU München, Holtzbrinck Publishing Group
Implementation period:01.01.2022 – 31.12.2023