Thanks for this post! The attention to detail in an easy to understand language was awesome.

Thank you for your post. What is a main advantage of neural process over gaussian process?
I think both models are for function approximation with uncertainty estimation and want to know the advantages of neural process. Thank you!

The conceptual difference is that GPs provide a prior over functions (which depends on the choice of kernel), whereas with NPs you first have to *learn* what sort of functions you want to capture. So NPs can be seen as a data-driven prior over functions (in principle this could be more flexible, but it means you need to (pre)train your NP before using it).

Thank you so much!!

Thanks for the post. But it looks like you only discussed the generation part, but not the inference part of NP?

Thank you very much for the clear and thorough explanation of neural Processes! I have ported your R code to Python and am creating a Jupyter notebook + presentation around it, here: https://github.com/sgvandij.... I'd like to include your diagrams, as I probably won't be able to make nicer ones :) but I wanted to check if that's ok with you first.

Thanks for the kind words, glad you found it useful! Sure, feel free to use my diagrams as long as you refer to the author.

Thank you so much for sharing!
It seems that the algorithm is very similar to VAE. Can I say the neural processes is a supervised learning version of VAE? If not, do they have any other difference?

I agree that there are some similarities between NPs and VAEs. I recommend you check out the discussion on it in the NP paper (see Figure 2 and Section 3.5).

Thank you so much for sharing your thoughts on this. This got me excited on NPs.