Abstract

Deep learning approaches have revolutionized artificial intelligence, but model opacity and fragility remain significant challenges. The reason for these challenges, we believe, is a knowledge gap at the heart of the field -- the lack of well-calibrated metrics quantifying the similarity of models obtained using different architectures, training strategies, different checkpoints, or under different random initializations. While several metrics have been proposed, they are poorly calibrated and susceptible to manipulations and confounding factors, as well as being computationally intensive when probed with a large and diverse set of test samples. We proposed an integration of chain normalization of weights and centered kernel alignment that, by focusing on weight similarity instead of activation similarity, overcomes most of the limitations of existing metrics. We are extending this approach towards complex model architectures and real-world chest X-ray image analysis. Moving forward, we aim to reveal how data, model architectures, hyperparameters, and training process shape the underlying "knowledge" learnt by deep neural network models.