@tcarpenter216 Gen.jl thesis is a great "computer science" take on building models as arbitrary programs and doing diverse things with them. The likelihood is very simple in this formulation. Great for "programmer brain" at least.
One word that English needs to borrow from Swedish is "shit." As in a mild exclaimation, like "darn," that a small child might say in the middle of a family Christmas special on TV.
@abecedarius Maybe it's time for me to do some tourism in Python-land. The guy who wrote the book on SMC is maintaining an interesting framework there (https://t.co/gPa04gq7EK). I'm also curious about JAX.
Ending this week on the thought: is Nested Sequential Monte Carlo sampling eerily similar to Monte Carlo Tree Search? Have to chase up how/if MCTS is applied in Bayesian context.
@abecedarius Gen.jl has an example where they call out to PyTorch to train a neural network to make proposals:
https://t.co/cgug7p2ytS
That sounded pretty out-there to me the first time I encountered it but maybe it's the common sense approach to good proposals in high dimensions.
@abecedarius Generally I'm warming up to the idea of black-box proposal distributions in Sequential Monte Carlo e.g. via deep learning, reinforcement learning, etc.
I mean, why not? Seems like (...) it should be straightforward to meet the requirements for Bayesian hygiene e.g. proper weight.