Using FHE(Fully Homomorphic Encryption) for machine learning.
Many schemes have different message space, which vary significantly from just binary bit of 0 or 1 to complex-valued matrices.
I think some of the operations in machine learning can be done more efficiently by packing the raw data into vectors, provided that it is fast enough to bootstrap and has somewhat more efficient way to do vector operations than just repeating additions and multiplications.
Blockchain networks based on these (quantum-safe?) FHE's are open to ML. For example, in an IoT environment, this blockchain network will make possible protecting user info but simultaneously giving personalized AI services.
But finding suitable format and scheme to pack data and setting the machine to do the learning process is still extremely tricky. Maybe we need more tests on real data to classify what should be adequate schemes and methods for a specific purpose in ML.