GoodAI (Czechia)

Designing faster optimization algorithms is of ever-growing interest. In recent years, learning to learn methods that learn how to optimize demonstrated very encouraging results. Current approaches usually do not effectively include the dynamics of the optimization process during training. They either omit it entirely or only implicitly assume the dynamics of an isolated parameter. In this paper, we show how to utilize the dynamic mode decomposition method for extracting informative features about optimization dynamics. By employing those features, we show that our learned optimizer generalizes much better to unseen optimization problems in short. The improved generalization is illustrated on multiple tasks where training the optimizer on one neural network generalizes to different architectures and distinct datasets.

Artificial Intelligence (AI) has an increasing impact on all areas of people's livelihoods. A detailed look at existing interdisciplinary and transdisciplinary metrics frameworks could bring new insights and enable practitioners to navigate the challenge of understanding and assessing the impact of Autonomous and Intelligent Systems (A/IS). There has been emerging consensus on fundamental ethical and rights-based AI principles proposed by scholars, governments, civil rights organizations, and technology companies. In order to move from principles to real-world implementation, we adopt a lens motivated by regulatory impact assessments and the well-being movement in public policy. Similar to public policy interventions, outcomes of AI systems implementation may have far-reaching complex impacts. In public policy, indicators are only part of a broader toolbox, as metrics inherently lead to gaming and dissolution of incentives and objectives. Similarly, in the case of A/IS, there's a need for a larger toolbox that allows for the iterative assessment of identified impacts, inclusion of new impacts in the analysis, and identification of emerging trade-offs. In this paper, we propose the practical application of an enhanced well-being impact assessment framework for A/IS that could be employed to address ethical and rights-based normative principles in AI. This process could enable a human-centered algorithmically-supported approach to the understanding of the impacts of AI systems. Finally, we propose a new testing infrastructure which would allow for governments, civil rights organizations, and others, to engage in cooperating with A/IS developers towards implementation of enhanced well-being impact assessments.