– DRAFT –
WebML CG Teleconference – 13 December 2018
13 December 2018
anssik: welcome to the first WebML CG telcon. We'll meet on a monthly cadence when there's an agenda.
<Rafael> n+ Rafael
anssik: subscribe to https://github.com/webmachinelearning/meetings/ for updates, also sent to public-webmachinelearning@w3.org
… I'm aware of a bunch of folks still in process of joining, so expect to have a small group today, also timezones are a challenge
… today want to discuss three things: 1) high vs. low level API proposed by Greg; 2) use cases review and gauge whether we have consensus to use these as a starting point for the API definition 3) WebNN proof-of-concept update since TPAC
<Yuan> How do I join? It asked for password, but I don't have one.
https://lists.w3.org/Archives/Member/internal-webmachinelearning/2018Nov/0000.html
anssik: the group's current charter has in scope a low-level API for inference. Greg raised an issue whether that's the right abstraction level, Greg please introduce the issue
<Yuan> Yes it asks for user name and password. What should I enter?
<gregwhitworth> gregwhitworth: To try and disambiguate the difference between low and high level - we can have an operator level API or load model API
<chrisheilmann> Chris Heilmann: I can see that most of the use cases will be sending data to a pre-trained model for web developers, not doing own training on device. The main thing I am worrying about is performance as in a client-side (browser) JavaScript context, we never can control the machines it runs on.
Gabe: unclear whether the higher level API would mean to communicate with a pre-trained model?
<chrisheilmann> Chris Heilmann: I'd love to be able to have a use case where I can run a model on device and request information from it by sending data to it. The benefits would be avoiding latency to a cloud service, and keeping my data on my device.
gregwhitworth: our view is there could be both a high level API for loading a pre-trained model, and a low level API that allows building a computational graph by common building blocks, operations
anssik: Tomoyuki and Ningxin have put together a proposal for use cases, split into "High-Level Use Cases" aka application-level use cases and "Low-Level Use Cases" aka API-level use cases proposing specific API features.
anssik: my expectation is everyone on the call has seen these use cases, so we are not going to read them aload here. Any specific issues that would benefit from discussion on this call
… I'd like to start a week-long Call for Consensus for adopting these use cases as a starting point for the API definition.
… Concrete feedback should be provided as comments in the PR.
gregwhitworth: the high level use cases align nicely with our internal discussion
… few low level use cases seem a bit of a stretch, but not going to push back on those, custom layer is a good one
anssik: silence means consent
anssik: those of you at TPAC saw a proof-of-concept WebNN implementation for doing real-time object detection, real-time human pose estimation, and image classification
ningxinhu: progress on WebNN since TPAC, this is a low level API we're investigating, based on Chromium. At TPAC we updated the group we have 9 ops supported, investigating ops support across platforms
… new op is fully connected, important for some models
… in native API mapping, enhanced MPS implementation with MPSNNGraph
… POC focused on performance, so collected perf date for ONNX.js on top of WebNN POC
<belem> https://github.com/intel/webml-polyfill/blob/master/docs/supported_ops.md
[Ningxin explaining ResNet-50 performance data]
ningxinhu: new use case investigation: video semantic segmentation
… outcome: good segmentation accuracy, performance not yet real-time
Rafael: which Windows APIs are used in the WebNN POC
<ningxinhu> ningxinhu: the current WebNN POC is using clDNN (https://github.com/intel/clDNN) for both Windows and Linux
anssik: An unofficial draft of a possible call for a participation to an unconfirmed W3C Workshop on Web & Machine Learning
[DRAFT] W3C Workshop on Web & Machine Learning
anssik: Next telcon scheduled for 10 Jan 2019, right after the holiday so might cancel. Suggested agenda topics please post to the list.
anssik: thank you for calling in, let's get the use cases finalized so that we can start 2019 with API definition
Succeeded: s/.. POC/... POC
Succeeded: s/TOPIC: Any Other Business//
Succeeded: s/7 Jan/10 Jan/