hongbin

Hi @willtebbutt , I have implemented a simple step function fitting example [here](https://github.com/HamletWantToCode/model_zoo), it use a feedforward neural network plus a GP with ARD kernel, I also write a binary...

I have run into a problem when trying to make Stheno's output data type to be `Float32` ( In `Flux`, the default data type for neural network parameters is `Float32`,...

Yeah, here are my ideas about implementation of Neural-Kernel-Network kernel: 1. NKN should be a subtype of Stheno's `Kernel` type, and should have `ew` & `pw` interfaces 2. it's construction...

> I wonder whether we could just directly use Flux's chain type here? I also consider using it, but there maybe some performance related problem, I'll write something first and...

The reason I don't include activation functions in the PR are: 1. Only limited type of functions are allowed, currently just polynomials with positive coefficients & `exp` function which can...