Not on current silicon, mostly because of the sensors though, neural networks have been used for many many years in the branch predictor(Since Piledriver at least), not particularly complex or computationally expensive in low layers, in fact their value is in how cheap, quick & low memory they are. Intel had working concepts on Itanium around the time the "fast-path" method was finalised (~2003) and could very well have been using them for longer(But this is arguably the most secretive part of the CPU).
The kind of sensor array used in Radeon VII is much much more useful(Actual transistor temps with high sample rates) but there's definitely room for refinement.