The AIDA paper contributes to the trustworthiness of deep learning models

The AIDA project consortium has published a paper at the 19th Symposium on Intelligent Data Analysis (IDA 2021), which took place online, between April 26 and 28.

The article is entitled “Partially Monotonic Learning for Neural Networks” and it shows that it is possible to encode monotonic relations between variables in complex neural networks without significantly losing model accuracy. For example, when increasing the price of some goods or services, while all other variables remaining equal, real world knowledge indicates that sales will drop (or at least not increase). Our contribution provides adjustable mechanisms to encode this type of knowledge in the form of monotonic relations, leading to models that are both accurate and trustworthy.

The authors of this paper are Joana Trindade, João Vinagre, Kelwin Fernandes, Nuno Paiva, Alípio Jorge, from INESC TEC.