Supporting Trusted Reuse of AI Models

With the widespread use of artificial intelligence (AI) models in physical and life sciences, imaging researchers can automate image-based measurements by reusing disseminated trained AI models. The challenges of reusing shared AI models include (a) trust in models’ performance due to insufficient information about shared AI models and black-box nature of the models, (b) insufficient size of domain-specific training datasets for retraining, and (c) computational resources needed to reproduce the work leading to the disseminated AI model. We will address these challenges by introducing (1) characteristics of AI models derived from optimization curves to be included in AI model cards and (2) traceable fingerprints of AI models to their training data.  The presentation will demonstrate the value of quantitative metrics for multi-purpose reuse of AI models, as well as the value of AI model simulations and measurements to establish traceability in non-linear models.