3Consorzio Interuniversitario Nazionale per l'Informatica, CINI, (Italy) The rising popularity of deep learning has brought to light a fundamental limitation of neural network architectures: they lack the ability to provide interpretable justifications for their decisions, making them unsuitable for contexts where human experts require transparent explanations [1]. This abstract summarizes a newly introduced comprehensive approach to Explainable Artificial Intelligence (XAI), which demonstrates how a deliberate design of neural networks produces a family of interpretable deep learning models known as Logic Explained Networks (LEN) [2]. LENs only necessitate human-understandable predicates as input concepts and ofer logic explanations of the output predictions via a set of First-Order Logic (FOL) formulas build on these predicates (see an example in Figure 1). A very interesting feature of this model is its versatility, indeed LENs can be applied in many use cases, including as interpretable classifiers or to explain another black-box model. In case of interpretable classification, some design choices, like learning criterion and parsimony index, allows to achieve state-of-the-art results in the prediction accuracy while gaining transparency on the model's decision process [3]. Concerning the learning paradigms, LENs can be successfully trained to learn and provide explanations both in supervised and unsupervised learning settings [2, 4].
Italy ∗Corresponding author. Explain! as a Python package on PyPI: https://pypi.org/project/torch-explain/ with an extensive documentation that is available on read at https: //pytorch-explain.readthedocs.io/en/latest/
of Black_foot_albatross Acknowledgments This work was supported by TAILOR and by HumanE-AI-Net, projects funded by EU Horizon 2020 research and innovation programme under GA No 952215 and No 952026, respectively.