Mark van der Wilk

The idea has been floating around. The paper you mention is really interesting, but it probably goes beyond what is possible in the current framework (from a quick skim it...

Softmax uses Monte Carlo to estimate the expected log likelihood, so you do always need Adam. The nice thing about RobustMax is that for the full dataset, you can use...

So I think this is cool, but rather low priority. We can already scale up multiple models just by adding more computation, e.g. running models in parallel on different machines....

That is very odd. Can't think of anything off the top of my head that could cause it though, given that you're evaluating a sequence of tf operations... For speed,...

That is strange. Unfortunately, I do not have time to go through your code. However, one thing that you may need to keep in mind when using such a small...

I think this is because of the global `jitter_level` setting. Numerical stability is a problem in variational methods, since current implementations* need to invert noise-free covariance matrices, which do not...

With fresh eyes, I have found the issue this morning. For interest, here is how to reproduce the problem I was having: ``` import numexpr as ne a = rnd.randn(1000,...

Apologies again, ``` import numexpr as ne a = rnd.randn(1000, 32)[:, None, :] a = np.asfortranarray(a) # I loaded a matrix from file, which (unknown to me) was Fortran contiguous...

Ok, the code is a bit complicated here, since there are quite a few cases. I suspect the error is in `Add.eKzxKxz()`. Your kernels work on separate dimensions and the...

We are in the process of refactoring the code at the moment. We will release an update with quite a few improvements to the code and numerical stability. I'm not...