Software development using free software licenses empowers us to use, modify, learn from, and build upon the work of others. Similarly, collaborative improvement of machine learning models using free software principles for training data, inference data, and machine learning models can empower development of tools we can all use, learn from, and understand.
But licenses that have worked so well for free software are now being applied – sometimes inconsistently – to the input data and results of machine learning. Does applying these licenses to data preserve software freedom?
In this talk, we explore this issue in three parts:
1. What does software freedom mean in a world where software is replaced by machine learning models?
2. Are there deficiencies in current machine learning licensing practices that we should address?
3. What is the legal impact of applying FOSS licenses to training data and machine learning models?