Models seem to be constructed by making them likely in the light of data but not such that they’re able to reproduce it. More ramblings about the likelihood.
An interesting statement: “neural networks use finitely many highly adaptive basis functions whereas gaussian processes typically use infinitely many fixed basis functions” - paraphrased from Wilson et al. 2015, based on work by MacKay, Neal and others.
Living a few months in Cambridge: (red is day, blue is night)
I love the blue mass at the astronomy centre and the movie theatre. More random projects here.