Multi-disciplinary Use Cases for Convergent new Approaches to AI explainability

开发和测试方法，允许从透明度、可解释性和可解释性方面解释人工智能算法的预测，已成为当今人工智能领域最重要的开放性问题之一。在本提案中，我们汇集了来自不同领域的研究人员，他们具有互补的技能，对于能够理解人工智能算法的行为至关重要，这将通过一组有趣的多学科用例进行研究，其中可解释的人工智能可以发挥关键作用，并将用于量化优势，突出和可能解决，在不同的应用环境中可用的可解释人工智能方法的弱点。迄今为止，阻碍可解释性取得实质性进展的一个方面是，在可解释性人工智能中，一些提出的解决方案在针对特定应用进行定制后被证明是有效的，而且往往不容易转移到其他领域。在这个项目中，我们将测试相同的技术阵列，以解释用例的可解释性，这些用例被故意选择为与数据类型、学习任务、科学问题相关的相当异构的用例。拟议的用例范围从高能物理人工智能应用，到应用于医学成像的人工智能，应用于肺、气管和鼻气道疾病诊断的人工智能，到用于改进神经科学分析和建模的可解释性机器学习技术。对于每个用例，研究项目将由三个阶段组成。在第一部分中，我们将根据需求适当选择最先进的可解释性技术应用于所考虑的案例。在第二部分，将指出这些技术的缺点。最值得注意的是，将考虑到高维数据和原始数据的可扩展性问题，其中噪声与感兴趣的信号相比可能普遍存在，只要每个算法提供的可验证性水平。在最后阶段，将每个用例中构建的算法和知识结合起来，以记录结果，并开发用于在一般和不同领域中利用xAI方法的通用程序和工程管道。

Developing and testing methodologies that allow to interpret the predictions of the AI algorithms in terms of transparency, interpretability, and explainability has become today one of the most important open questions in AI. In this proposal we bring together researchers from different fields with complementary skills, essential to be able to understand the behaviour of the AI algorithms, that will be studied with an interesting set of multidisciplinary use-cases in which explainable AI can play a crucial role and that will be used to quantify strengths and highlight, and possible solve, weaknesses of the available explainable AI methods in different applicative contexts. One aspect hindering so far substantial progress towards explainability is the fact that several proposed solutions in explainable AI proved to be effective after being tailored to specific applications, and frequently not easily transferred to other domains. In this project, we will test the same array of techniques for explainability to use-cases intentionally chosen to be quite heterogeneous with respect to the types of data, learning tasks, scientific questions. The proposed use-cases range from High Energy Physics AI applications, to applied AI in medical imaging, to AI applied for the diagnosis of pulmonary, tracheal and nasal airways diseases, to machine-learning techniques of explainability used to improve analysis and modelling in neuroscience. For each use-case, the research project will consist of three phases. In the first part, we will apply state-of-the-art explainability techniques, properly chosen based on the requirements, to the case under consideration. In the second part, shortcomings of the techniques will be identified. Most notably, issues of scalability to high-dimensional and raw data, where noise can be prevalent compared to the signal of interest, will be taken into consideration, as long as the level of certifiability afforded by each algorithm. In the final phase, algorithms and knowledge built in each use-case will be combined in order to document the results and to develop general procedures and engineering pipelines useful for the exploitation of xAI methods in general and different domains.