Towards interpretable-by-design deep learning algorithms
This addresses the issue of explainability in deep learning for users needing transparent AI systems, though it is incremental as it builds on existing foundation models.
The paper tackles the problem of making deep learning models interpretable by proposing the IDEAL framework, which recasts classification as similarity to prototypes derived from training data while leveraging pre-trained foundation models, resulting in interpretable models that mitigate confounded interpretations, avoid catastrophic forgetting, and enable efficient transfer learning without fine-tuning.
The proposed framework named IDEAL (Interpretable-by-design DEep learning ALgorithms) recasts the standard supervised classification problem into a function of similarity to a set of prototypes derived from the training data, while taking advantage of existing latent spaces of large neural networks forming so-called Foundation Models (FM). This addresses the issue of explainability (stage B) while retaining the benefits from the tremendous achievements offered by DL models (e.g., visual transformers, ViT) pre-trained on huge data sets such as IG-3.6B + ImageNet-1K or LVD-142M (stage A). We show that one can turn such DL models into conceptually simpler, explainable-through-prototypes ones. The key findings can be summarized as follows: (1) the proposed models are interpretable through prototypes, mitigating the issue of confounded interpretations, (2) the proposed IDEAL framework circumvents the issue of catastrophic forgetting allowing efficient class-incremental learning, and (3) the proposed IDEAL approach demonstrates that ViT architectures narrow the gap between finetuned and non-finetuned models allowing for transfer learning in a fraction of time \textbf{without} finetuning of the feature space on a target dataset with iterative supervised methods.