Hi there,

I continue to learn Turing. In the tutorial for Bayesian Logistic Regression (https://turing.ml/dev/tutorials/2-logisticregression/), there is a section to check the trend plots of the chains, in order to “do a spot check to make sure each chain converges around similar points”. However, in the tutorial for Bayesian Neural Network (https://turing.ml/dev/tutorials/3-bayesnn/), there is not such a section to check the convergence of chains. I sampled three chains using

mapreduce(c -> sample(bayes_nn(hcat(xs…),ts), HMC(0.05, 4), 1500),

chainscat,

1:3

)

The chains do not converge according to a trend plot.

Would you please help me understand to how to make the chains converge in a BNN model? I understand that I can use “MAP estimation to classify our population”, but I prefer to use the average values of chains as parameter estimates.

Thanks,

Chuan