Deep Learning (DL) models excel at automatically learning intricate patterns within complex data, but their black box nature undermines human trust. To address this, current validation strategies typically focus on the model itself, modifying its architecture to assess the role and importance of the components. However, this model-centric view overlooks the critical learning substrate, which is represented by the data, implicitly assuming that it accurately represents the target phenomenon. This implicit trust in data means that evaluation may fail to detect whether high performance stems from exploiting biases or data quirks rather than learning relevant patterns. We present a novel data-related ablation as a complement to the traditional architectural ablation. Using this framework for Electroencephalography (EEG) signals of Emotional Recognition (ER) and Motor Execution (ME) as a case study, we show that seemingly high-accuracy models often rely heavily on process-irrelevant features, maintaining performance even when key information is eliminated. This shows that a standard, data-independent evaluation can be misleading about whether a model truly captured the intended process; the proposed approach helps distinguish robust learning from leaning on incidental characteristics. Therefore, incorporating data-related ablation is essential for developing reliable and generalizable DL models in fields that rely on data derived from complex and often not completely known phenomena.
Lanzino et al. (Fri,) studied this question.