I'm Dan Forman Mackey.

I'm an associate research scientist here at CCA.

I sit on the fifth floor.

I'm glad I get to go first so I can set the bar appropriately low.

So I wanted to spend this time to introduce myself and talk about some of the projects that I'm working on and that I'm hoping to collaborate on.

And I put my introduction slide as my GitHub profile because that's where I keep track of all the work that I do.

Does PyMC3, does that use PyTorch for derivatives, or how do you compute the derivative?

Right, so it's built on top of Theano, which is kind of a precursor for those kinds of things.

Do you run into issues with like multimodal parameters where maybe one Walker is stuck in one and other walkers and another and if they try to move to each other, they end up spending time in the middle, or is that might not be realistic.

Great so the question was what happens if you have a multimodal distribution you your helper Walker is in a different mode than the Walker that you're that you're moving.

And that is bad news right because you're still doing an accept reject step that is this is similar to the metropolis hastings except reject and so, if you're let's say we have a one D. posterior that kind of looks like this, and our one walker is here and our other walker is over here, your proposal is mostly going to be in regions that are very low probability.

So you're not going to accept them.

And so it doesn't make the algorithm wrong.

It's not going to spend more time than it should in between the modes.

But your acceptance fraction is going to plummet.

Sure.