This case study thus investigates potential strategies and methods to expose the epistemic failures of AI systems in the natural and social sciences, with a particular focus on epistemic/inductive biases in pre-trained deep learning models.

As protein folding has been claimed to be the one “win” for AI research that’s not just decorative (i.e. generative) we decided to look at this first. The models in this domain are also based on the transformer architecture, which underlies almost all recent research, and is the current challenge of interpretability research. What we are especially interested is how the language paradigm in biology (e.g. “genetic code”) and the (operationalized-as-transformer) language paradigm in machine learning are mapped to each other/historically linked.