Kevin P Murphy

replay (iterated) LOFI

1

Another way to avoid overcounting would be to use the buffer for updating the linearization point without updating mu or sigma. For example, at the beginning of step t we...

See sec E.6" "Learning from the loss gradient"

Translate this theano/pymc code into blackjax https://github.com/TeaPearce/Expressive_Priors_in_BNNs/blob/master/code/timeseries_hmc_05.py This should reproduce fig 3 from te paper below (first 3 columns, ignore the GP part) T. Pearce, R. Tsuchida, M. Zaki, A....

order of arguments to build_xxx_kernel is arguably confusing (stylistic issue)

1

The [API for build_xxx_kernel](https://github.com/jeremiecoullon/SGMCMCJax/blob/master/sgmcmcjax/kernels.py) has the form ``` build_kernel(varargs, loglikelihood, logprior, data, batch_size, kwargs) ``` where varargs is a variable number of arguments, depending on the kernel. For example ```...

- In examples/gmrf.ipynb, maybe replace `from jax.example_libraries import optimizers` with optax. Also the term "GMRF" usually refers to Gaussian MRF, not grid :) Finally the learning code is super slow,...

I added a very short new section on training objectives for deep generative models, going beyond the likelihood maximization approach adopted in most of the book. This is motivated by...

feature request: add support for accessing log normalizer

9

Computing the log parition function for a conjugate family is a very useful quantity eg for computing the marginal likelihood p(D), which is needed for empirical Bayes and model selection....

see https://github.com/probml/pml-book/issues/670

see https://github.com/probml/pml-book/issues/671

Objective should be max \sum_i ||z_i||^2 s.t. || z_i - z_j || = || x_i - x_j || for neighbors. This would give max Tr(K) s.t. K_ii _+ K_jj -...