Abstracts

HUMAN-CENTRIC EXPLAINABLE AI FOR NEUROSCIENCE: REVEALING NEURON POPULATION ACTIVITY

Davor Horvatić1, Domjan Barić1, Petar Fumić1, Tomislav Lipic 2

1Department of Physics, Faculty of Science, University of Zagreb, Zagreb, Croatia; 2Division of Electronics, Ruđer Bošković Institute, Zagreb, Croatia

Despite the outstanding advancements and potential benefits of deep learning-based AI solutions, their black-box nature and the lack of transparency limits them as a tool for scientific discovery. To fully trust, accept, and adopt newly emerging AI solutions, we need human-centric explainable AI (HC-XAI) to provide human-understandable interpretations for algorithmic behaviour and outcomes. This presentation will focus on interpretable deep learning models for multivariate forecasting tasks. Several architectures will be presented with their performance score and correct interpretability tested on a novel benchmark of synthetically designed datasets with the transparent underlying generating process of multiple time series interactions with increasing complexity. I.e., one wants to test if these architectures can capture correct autocorrelations and crosscorrelations between multiple time series. This benchmark enables us an empirical evaluation of the performance of deep neural networks in three different aspects: prediction performance score, interpretability correctness, and sensitivity analysis. Furthermore, the synthetic approach enables us to build a good foundation for understanding neural population activity in measurements that significantly range in population size scale.

Acknowledgments: D.H. and F.P. were funded by QuantiX—Lie Center of Excellence, a project co-financed by the Croatian Government and European Union through the European Regional Development Fund—The Competitiveness and Cohesion Operational Programme (grant KK.01.1.1.01.0004). T.L. was funded by the Centre of Excellence project “DATACROSS”, co-financed by the Croatian Government and the European Union through the European Regional Development Fund—the Competitiveness and Cohesion Operational Programme (KK.01.1.1.01.0009).

Presenting author e-mail address: davorh@phy.hr