Trusting in the Sciences Requires Explainability

The July issue of IEEE Signal Processing Magazine (SPM) is a special issue focused on “Explainability in Data Science: Interpretability, Reproducibility, and Replicability.” With increased enthusiasm for machine learning, it is a very timely topic, and I invite every IEEE Signal Processing Society (SPS) member to read these very instructive papers.

The “classical” scientific approach is based on models, designed with realistic assumptions on the problem to solve. The main advantage of such an approach is to provide direct explainability of the results. The limitation is the model’s complexity, which is limited: it is usual to claim following G. Box: “ a model is always wrong, but it can be useful” [1].

With the data deluge we have experienced in recent years, many more data-driven methods have been developed with great success. Their main advantage is that they do not first require the designing of a model. This is also their drawback as many of them are black boxes: explaining their results is tricky and requires a lot of effort and additional experiments. As data science, including machine learning and deep learning, are currently ubiquitous in all domains, explainability in data science is essential, especially for critical application domains like medicine and health, control of autonomous vehicles, and face recognition, to name a few.

In addition to intellectual satisfaction and trust for the user provided by explainability, in the European Union, as it is recalled in the article “Robust Explainability” in this month’s issue (page 73), the General Data Protection Regulation adopted in May 2018 states that individuals have the right to an explanation of a decision based on automated processing [2].