Machine Learning: Challenges for FOSS


What does it mean for a trained neural network to be Free Software? What needs to be shared to allow users to exercise their freedom to adapt and improve it? Do our current licences fit the task?

More and more software-mediated tasks are being accomplished using trained neural networks, in part because the hardware that facilitates deep learning has become so much faster and cheaper. How does the FOSS licensing and development model interact with this trend? Challenges include:

+ A lack of clarity around the extent and character of IP present in a trained neural network
+ A reproducibility problem around the generation of a trained neural network from its ‘source’ data and parameters
+ Reliance on closed hardware drivers to achieve realistic training rates

At a time when our personal freedoms rely more and more on decisions made by such systems, are traditional models of software freedom failing to protect us?



Supported by