Re: WebML CG Teleconference - 14 February - call for agenda topics

Hello all,

Probably won’t be able to make the call, so leaving feedback here (and on GH):

#3 eager execution and #6 are sort of related. Both touch on opening up a bunch of math primitives to the web which is great, but also hard. While I personally would love to see this happen, scoping the operations and tradeoffs is really hard; and improvements on the syntactic infrastructure (since you definitely want certain operations to be infix for ergonomic reasons) needs to be delivered from another SDO. (TC39, in particular)

This will take a lot of coordination, so I think clearly splitting the scope between “what we can do now” and “what we eventually want to do” to have everyone aligned on what the group is planning to deliver.
 
Cheers,
Sangwhan Moon

> On Feb 8, 2019, at 0:07, Kostiainen, Anssi <anssi.kostiainen@intel.com> wrote:
> 
> Hi All,
> 
> We're slotted to have our monthly teleconference next Thursday, 14 February, 14:30-15:00 UTC+0. I observe two issues #3 and #6 on the table we could have a good discussion on. Before sending the logistics I'd like to check if this makes for a good agenda to you all.
> 
> Please send any suggestions by EOB Tuesday, 12 February.
> 
> Since we have only 30 mins, any in-depth input should be provided via webnn repo's GH issues ahead the call for others to digest.
> 
> Proposed agenda
> 
> #3 High level vs low level
> https://github.com/webmachinelearning/webnn/issues/3
> 
> We've made progress on the API abstraction level issue, but no consensus has emerged yet. Latest investigation is on mapping native APIs' eager vs. graph execution modes to corresponding Web APIs, and implications thereof.
> 
> #6 Custom operations
> https://github.com/webmachinelearning/webnn/issues/6
> 
> I observe positive signals. Discuss concrete steps on how we can move forward with this investigation as a group. Possible tasks include e.g. looking into technical details on how browsers plan to implement built-in ops, survey what ops would be typically needed to write custom ops.
> 
> (Happy Chinese New Year 2019 to our Chinese participants!)
> 
> Thanks,
> 
> -Anssi (WebML CG chair)

Received on Friday, 8 February 2019 03:18:20 UTC