Back in 2010, during my PhD, I explored some ideas for learning twist functions for SMC. (The twists were linear random feature models since this was pre-DL-era.) I didn't try to publish since I couldn't think of a compelling use case. Sometimes you just have to wait.