There are existing software implementations, particularly the various free-energy principle implementations available in SPM from UCL (albeit, in Matlab), and a

There are existing software implementations, particularly the various free-energy principle implementations available in SPM from UCL (albeit, in Matlab), and a few others. What has tended to happen is that neuroscientists are more interested in a minimal bio-plausible proof of principle, and machine learners often don't keep entirely up to date with neuroscience after their training. As a result, a lot of the most interesting ideas in neuroscience haven't been scaled up to larger machine-learning applications yet. Me being a dual trainee in both fields, well, I'm working on it.