Sónia Carvalho Teixeira

Abstract
This Portuguese project compares the classification of AI risks and vulnerabilities performed by humans and performed by the computing algorithms.

Abstract
Our lives have been increasingly filled with technologies that use Artificial Intelligence (AI), whether at home, in public spaces, in social organizations, or in services. Like other technologies, adopting this emerging technology also requires society's attention to the challenges that may arise from it. The media brought to the public some unexpected results from using these technologies, for example, the unfairness case in the COMPAS system. It became more evident that these technologies can have unintended consequences. In particular, in the public interest domain, these unintended consequences and their origin are a challenge for public policies, governance, and responsible AI. This work aims to identify the technological and ethical risks in data-driven decision systems based on AI and conduct a diagnosis of these risks and their perception. To do that, we use a triangulation of methods. In the first stage, a search on Web of Science has been performed. We consider all the 412 papers. The second stage corresponds to a analysis of experts. The papers have been classified according to the relevance to the topic by the experts. In the third stage, we use the survey method and include risk insights from stage two in our questions. We found 24 concerns which arise from the perspective of the ethical and technological risk perspective. The perception of participants regarding the level of concern they have with the risks of a data-driven system based on AI is high than their perception of society's concern. Fairness is considered the risk whose perception is more severe. Fairness, Bias, Accountability, Interpretability, and Explainability are considered the most relevant concepts for a responsible AI. Consequently, also the most relevant for responsible governance of AI. © 2022 ACM.

Abstract

Abstract
We present an online optimization method for time-evolving data streams that can automatically adapt the hyper-parameters of an embedding model. More specifically, we employ the Nelder-Mead algorithm, which uses a set of heuristics to produce and exploit several potentially good configurations, from which the best one is selected and deployed. This step is repeated whenever the distribution of the data is changing. We evaluate our approach on streams of real-world as well as synthetic data, where the latter is generated in such way that its characteristics change over time (concept drift). Overall, we achieve good performance in terms of accuracy compared to state-of-the-art AutoML techniques.

Abstract