For discussing how to improve the BatchNorm section. I am working on a draft PR.

Currently there is a good note/example, and description of the parameters.

Formal steps are missing. They should spec how the input and output is handled. For instance, whether normalization can be done in-situ. May the impl modify the input and return the same object (I assume not, i.e. different JS objects, but may share underlying buffers).

In addition, I thought to add the batch norm algorithm from the linked paper as an informal, descriptive sub-steps section.

Also, somewhat more detailed introduction, including typical use cases (currently shortly mentioned).

Any inputs, examples, discussion welcome.