Benchmarking and Enhancing Reliable AI: Knowledge-Grounded Architectures and Alternative Optimization Mechanisms

AI systems are pervasively providing solutions in several application domains, in order to solve different types of problems. This scenario raises ethical concerns regarding the fairness of the inference process and the verifiability of the result. Furthermore, such concerns is also bounded to the sustainability of large solutions and their significant impact on environmental resources. The objective of this research project is to contribute to the resolution of these issues. A preliminary goal is to establish equitable benchmarking frameworks, on diverse desiderata qualities of trustworthy AI. A possible solution we investigate is the integration of structured knowledge in a neuro-symbolical manner and probabilistic calculus, to ground the learning stages. Then, further investigations include the assessment of the integration of backpropagation alternatives and biological-inspired computation for optimization purposes. Accordingly, this paper delineates the analytical process that led to the formulation of the research project’s conceptual framework and the possible domains in which such techniques could be applied.