Dr. Abeba Birhane:

I just want to clarify what we mean by black box technology or black box algorithms. It generally refers to the fact that, for example, for a computer vision system or a face recognition system you save up your parameters, you gather up thousands and even millions and billions of images of faces, and you train your algorithm to differentiate what is a face from what is not a face. After sufficient training, then the algorithm has learned what a face is like. However, you do not know how. That is what is generally referred to as a black box algorithm. This at the heart of machine learning. It is a problem not only with face recognition systems or computer vision but AI in general.

However, it is still possible to have as much information about a given algorithm to be able to assess error rates, performance and so on. For example, training data is a critical element that determines how a certain algorithm performs. Things such as training data and the model architecture are what I refer to when I talk about the public or independent auditors - not even the public, as they do not have to give access to the public. This is information that is available which independent auditors could have access to and could do independent evaluations of such given models. This is the kind of information that is hidden by proprietary rights that nobody outside the vendors and probably An Garda Síochána - I am not sure about the arrangements - may have access to. I just wanted to clarify what we mean by a black box algorithm.