08:57:24 <RRSAgent> RRSAgent has joined #webmachinelearning
08:57:24 <RRSAgent> logging to https://www.w3.org/2018/10/24-webmachinelearning-irc
08:57:29 <Guru> Guru has joined #webmachinelearning
08:57:31 <Zakim> Zakim has joined #webmachinelearning
08:57:32 <anssik> RRSAgent, make logs public
08:57:37 <anssik> Meeting: Machine Learning for the Web breakout at TPAC 2018
08:58:12 <anssik> RRSAgent, draft minutes v2
08:58:12 <RRSAgent> I have made the request to generate https://www.w3.org/2018/10/24-webmachinelearning-minutes.html anssik
08:58:13 <ningxinhu> ningxinhu has joined #webmachinelearning
08:58:21 <ningxinhu> Present+ NingxinHu
08:58:39 <anssik> RRSAgent, make logs public
08:59:45 <fabien_gandon> fabien_gandon has joined #webmachinelearning
09:00:00 <myles_> myles_ has joined #webmachinelearning
09:00:03 <fabien_gandon> present+ Fabien Gandon
09:00:14 <johnchen> johnchen has joined #webmachinelearning
09:00:17 <kajimoto_> kajimoto_ has joined #webmachinelearning
09:00:23 <myles_> Present+ myles_
09:00:29 <anssik> Present+ Anssi_Kostiainen
09:00:36 <anssik> -> https://www.w3.org/community/webmachinelearning/ Machine Learning for the Web Community Group
09:00:40 <kajimoto_> Present+ kajimoto
09:00:54 <anssik> -> https://webmachinelearning.github.io/charter/ Machine Learning for the Web Community Group Charter
09:01:38 <tidoust> tidoust has joined #webmachinelearning
09:01:50 <tidoust> Present+ Francois_Daoust
09:01:57 <tidoust> scribenick: tidoust
09:02:23 <halo> halo has joined #webmachinelearning
09:02:24 <anssik> anssik has changed the topic to: https://www.w3.org/wiki/TPAC/2018/SessionIdeas#Machine_Learning_for_the_Web
09:03:11 <Kangz> Kangz has joined #webmachinelearning
09:03:17 <anssik> Chair: Anssi
09:03:19 <Kangz> Hi, this is Corentin Wallez
09:03:22 <anssik> RRSAgent, draft minutes v2
09:03:22 <RRSAgent> I have made the request to generate https://www.w3.org/2018/10/24-webmachinelearning-minutes.html anssik
09:03:38 <Huangweichai> Huangweichai has joined #webmachinelearning
09:03:38 <Yanni> Yanni has joined #webmachinelearning
09:03:53 <tidoust> Agenda: https://www.w3.org/wiki/TPAC/2018/SessionIdeas#Machine_Learning_for_the_Web
09:03:55 <hirata> hirata has joined #webmachinelearning
09:04:19 <sangchul> sangchul has joined #webmachinelearning
09:04:33 <Barbara_Intel> Barbara_Intel has joined #webmachinelearning
09:04:41 <dom> dom has joined #webmachinelearning
09:04:57 <huangweichai_> huangweichai_ has joined #webmachinelearning
09:05:16 <tidoust> Topic: Introduction
09:05:28 <sangwhan> sangwhan has joined #webmachinelearning
09:06:11 <iank_> iank_ has joined #webmachinelearning
09:06:27 <tidoust> anssik: Welcome to the Machine Learning for the Web breakout session at TPAC! I'll be giving a quick background on this idea.
09:06:36 <ktoumura> ktoumura has joined #webmachinelearning
09:06:42 <Lin> Lin has joined #webmachinelearning
09:06:47 <zolkis> zolkis has joined #webmachinelearning
09:06:51 <Youngsun> Youngsun has joined #webmachinelearning
09:07:04 <tidoust> ... We proposed and created a Community Group a few weeks ago. Before that, we did some incubation in the Web Platform Incubator Community Group (WICG)
09:07:04 <zolkis> present+ Zoltan_Kis
09:07:06 <jungkees> jungkees has joined #webmachinelearning
09:07:08 <ktoumura> present +Kunihiko_Toumura
09:07:08 <hyojin> hyojin has joined #webmachinelearning
09:07:11 <tidoust> ... This helped set the scope of the work.
09:07:29 <dom> Present+ Dominique_Hazael-Massieux
09:07:40 <tidoust> ... Basically, the work is scoped to neural networks. Low-level API that can be implemented accross platforms to do client-side inference.
09:07:48 <dom> RRSAgent, draft minutes
09:07:48 <RRSAgent> I have made the request to generate https://www.w3.org/2018/10/24-webmachinelearning-minutes.html dom
09:07:50 <tidoust> ... Why do we want to do that?
09:08:28 <MichaelGordon> MichaelGordon has joined #webmachinelearning
09:08:29 <tidoust> ... Getting closer to the platforms improves performance. For object recognition and immersive computation, that's crucial.
09:08:47 <tidoust> ... But there are also many computation needs that you need to do on the client.
09:08:53 <GeunHyung> GeunHyung has joined #webmachinelearning
09:09:00 <tidoust> ... I would like to start with a demo.
09:09:03 <MichaelGordon> Present+ MichaelGordon
09:09:14 <yang_gu> yang_gu has joined #webmachinelearning
09:09:47 <ningxinhu> https://huningxin.github.io/webml-examples/
09:09:54 <sangwhan> present+
09:10:08 <tidoust> ... The demo works on your Web browsers right now because we're using a polyfill built on top of WebGL, and later WebGPU.
09:10:26 <tidoust> ningxinhu: [going through demo]
09:10:36 <Yaohua_Wu> Yaohua_Wu has joined #webmachinelearning
09:10:39 <Guru> Guru has joined #webmachinelearning
09:10:49 <tidoust> ... Model gets downloaded first. In this case, cached through Service Worker.
09:11:20 <tidoust> ... The demo uses Koko(sp?) that can recognize about 100 objects.
09:11:27 <jmann> jmann has joined #webmachinelearning
09:11:31 <tidoust> ... s/Koko(sp?)/Coco/
09:11:38 <jmann> Present+ Jatinder Mann
09:12:16 <tidoust> ... The WebML propotype uses Chromium. Two polyfill implementations: WebAssembly and WebGL.
09:12:42 <tidoust> ... WebAssembly is much slower, because WebAssembly lacks SIMD features that allow to parallelize computation.
09:12:54 <sangchul_> sangchul_ has joined #webmachinelearning
09:13:17 <tidoust> ... WebGL2 is faster. WebML is the fatest.
09:13:34 <tidoust> anssik: The prototype is quite convincing in terms of performance gains with WebML.
09:13:49 <tidoust> ... We have a F2F meeting of the CG this Friday afternoon.
09:13:59 <tidoust> ... We'd like to hear from your use cases to start with.
09:14:30 <tidoust> ... The primary user of the current API is authors of frameworks, e.g. TensorFlow.js
09:14:54 <tidoust> Topic: Overview of API
09:15:14 <lin> lin has joined #webmachinelearning
09:15:34 <tidoust> ningxinhu: The scope is that we want to do Edge AI computing.
09:15:40 <sangchul_> present+ Sangchul_Ahn
09:15:44 <lin> present+
09:15:46 <tidoust> ... [showing presentation slides]
09:16:00 <tidoust> ... Sentiment anaylsis, etc.
09:16:25 <sangwhan> q+ to suggest better naming, ask about rationale of the "whitelisted" layers, choice of integer constants over string enums, choice of procedural programming practices for model definition, output layer being re-used being somewhat non-idiomatic javascript, and more
09:16:29 <anssik> RRSAgent, draft minutes v2
09:16:29 <RRSAgent> I have made the request to generate https://www.w3.org/2018/10/24-webmachinelearning-minutes.html anssik
09:16:44 <tidoust> ... JS Machine Learning frameworks such as TensorFlow.js, Keras.js, OpenCV.js provide the capabilities to do the inference on the client.
09:17:00 <tidoust> ... The models you've trained can be reused in different frameworks.
09:17:45 <tidoust> ... Some advantages to edge AI: latency, availability, privacy, and also cost as processing is delegated to the client. Users use their computation capabilities.
09:18:29 <tidoust> ... Gain of performance is also on the power consumption side. Current JS-based solutions consume a lot of power, especially on mobile devices.
09:18:59 <tidoust> ... We prepared some benchmark to compare performances.
09:19:39 <tidoust> ... I will not go into much details about the data, but the takeaway is that the gap between a Web-based and native solution is huge.
09:20:18 <tidoust> ... Even comparing a solution that uses WebGPU and a WebML approach, there's still a 4 times gain.
09:21:28 <tidoust> ... Looking at existing solutions today, CPU Deep Learning (DL) optimizations are not exposed to applications. That's crucial, in particular on mobile devices.
09:21:49 <tidoust> ... Also dedicated DL hardware accelerators.
09:22:24 <tidoust> ?1: Some use cases where deep learning is not the right approach. Why don't you try to support other kinds of algorithms?
09:23:09 <tidoust> ... Use case around high-res images
09:23:44 <tidoust> ningxinhu: My understanding is that DL needs low-res images, because training models would not work too well with high-res images.
09:24:09 <tidoust> anssik: We wanted to be practical with the scope. Minimal approach of what would enable use cases fast.
09:24:14 <sangchul_> sangchul_ has joined #webmachinelearning
09:24:47 <Yanni> Yanni has joined #webmachinelearning
09:24:52 <tidoust> ?1: There are use cases where you don't have all of the data to train the models, and Web developers will run into the issue.
09:25:33 <tomoyuki> s/KoKo(sp?)/COCO/
09:25:35 <lin> lin has joined #webmachinelearning
09:25:53 <tidoust> ningxinhu: Web developers won't have much data to train models, that's your concern?
09:25:57 <tidoust> ?1: Yes
09:26:34 <tidoust> ningxinhu: Web developers will be able to reuse other trained models and re-train them with additional data.
09:27:10 <tidoust> ... Here, our focus is how to tackle the performance issue for the Web. With that, the Web developer can leverage existing trained models.
09:27:26 <tidoust> anssik: It should be made clear that training is considered out of scope of this work.
09:27:29 <dom> q?
09:28:45 <tidoust> ningxinhu: Our proposing Web Neural Network (WebNN) is meant to leverage existing hardware infrastructure. The OS should ship their own Machine Learning API, abstracted away by WebNN.
09:29:29 <tidoust> ... WebNN provides low-level interfaces to these APIs to support different models, e.g. ONNX models, TensorFlow models, etc.
09:30:47 <tidoust> ... Positioning the API on the Web. A Usage API uses a built-in model, easy to ingrate. Maps to different platform primitives, such as NLP, Skills, MLKit.
09:31:27 <tidoust> ... The Model-level API (WebML) would map to CoreML, WinML, TensorFlow Lite, etc.
09:31:50 <Huangweichai> Huangweichai has joined #webmachinelearning
09:32:04 <tidoust> ... WebNN is the acceleration API, close to hardware optimization. Maps to BNNS, MPS, DirectML, Android NN.
09:32:15 <Yanni> Yanni has joined #webmachinelearning
09:32:16 <tidoust> ... WebNN is our current focus.
09:32:53 <tidoust> anssik: Right, the low-level WebNN proposal is the goal of the incubation in the Community Group. We're hoping this will lead to the standardization of the other higher-level APIs in the future.
09:33:17 <tidoust> Alex: WebNN is JS functions to manipulate models, is that correct?
09:33:29 <tidoust> ningxinhu: We can review the exact design afterwards.
09:34:36 <tidoust> anssik: You can refer to the following scope statement
09:34:37 <anssik> -> https://webmachinelearning.github.io/charter/ Machine Learning for the Web Community Group Charter
09:34:37 <dom> https://webmachinelearning.github.io/charter/
09:35:32 <tidoust> ningxinhu: On the top of this, WebML (not a good name), if you have interfaces for models, it will ease reuse of models. But that API can be built as a polyfill on top of the WebNN API in the near future.
09:36:12 <tidoust> ?2: [performance figures question]
09:36:32 <tidoust> ningxinhu: WebNN is within 10-20% of pure native performances.
09:36:58 <anssik> [POC = proof of concept]
09:37:11 <tidoust> ... We want to use a Proof of Concept as starting point and explore cross-platform capability.
09:37:45 <tidoust> ... The polyfill API is modeled from Android NNAPI.
09:38:34 <tidoust> ... You need to manipulate models, with operations and operands. Also, you need a compilation stage to initialize the hardware.
09:38:56 <tidoust> ... Last the execution stage where you process inputs according to the models and make inferance to produce outputs.
09:40:01 <tidoust> ... We're using a promise-based API, so when your promise is resolved, you have the answer.
09:40:26 <tidoust> anssik: If people want to see more demos, they can haz them!
09:41:29 <tidoust> ningxinhu: We started with a polyfill with WebAssembly and WebGL backends. Compute shaders.
09:42:44 <tidoust> ... and then developed a Chromium prototype, starting with MPS/BNNS API for MacOS, NN API for Android, cross-platform hardware agnostic.
09:42:56 <dom> q?
09:42:58 <tidoust> ... WebNN can leverage CPUs, GPUs and accelerators.
09:43:36 <GeunHyung> GeunHyung has joined #webmachinelearning
09:44:14 <tidoust> Myles: Wondering about numbers. What happens that explains the 24x times gap between CoreML (CPU) and WebAssembly implementations.
09:44:56 <tidoust> ningxinhu: Some missing threading and SIMD features explain the difference. The gap should be reduced once features are in.
09:45:34 <tidoust> Iank: The number is a bit unfair for WASM, because these features are incoming.
09:46:04 <tidoust> Myles: Let's pretend these features are used. How far down would that gap become?
09:46:28 <tidoust> ningxinhu: We'd probably go down to 4x.
09:47:00 <tidoust> sangwhan: Comparison should go with TensorFlow. Nested loops to do convolution would be hard to achieve.
09:47:28 <tidoust> Deepti: A lot of feedback about WASM SIMD is that we may not be covering the right set of instructions.
09:47:45 <tidoust> ningxinhu: Multiple add.
09:47:51 <dom> ack sangwhan
09:47:51 <Zakim> sangwhan, you wanted to suggest better naming, ask about rationale of the "whitelisted" layers, choice of integer constants over string enums, choice of procedural programming
09:47:54 <Zakim> ... practices for model definition, output layer being re-used being somewhat non-idiomatic javascript, and more
09:48:10 <lin> lin has joined #webmachinelearning
09:48:54 <tidoust> sangwhan: Question about the serialization of models, saw that in scope in the draft charter.
09:49:08 <tidoust> [draft charter shown]
09:49:41 <tidoust> sangwhan: To make interoperability less plainful, it would be good to include that in the scope.
09:50:33 <tidoust> ... Worried about the bytes in the array buffer.
09:50:37 <tidoust> anssik: I think you might want to open an issue on the GitHub repo
09:50:55 <zolkis> zolkis has joined #webmachinelearning
09:50:58 <tidoust> Corentin: There's a standard in the Khronos group for that.
09:51:21 <Kangz> https://www.khronos.org/nnef
09:51:25 <tidoust> sangwhan: Right, I just want to clarify what's in scope. Doesn't matter if it reuses an existing spec.
09:52:41 <Yanni> Yanni has joined #webmachinelearning
09:52:41 <tidoust> ningxinhu: WebNN is very low-level API. Main target is JS frameworks. They can add their logic to deal with such issues.
09:52:50 <tidoust> sangwhan: Not entirely convinced, but OK.
09:53:29 <tidoust> anssik: Feel free to influence the scope on GitHub. The draft charter is a starting point. We want to get your input. Not an easy task to land on a scope that seems to work for people.
09:53:42 <sangwhan> q-
09:54:06 <tidoust> Topic: Machine Learning for the Web Community Group
09:54:13 <anssik> -> https://github.com/webmachinelearning/meetings/blob/master/2018-10-26-lyon/README.md Machine Learning for the Web Community Group F2F at TPAC 2018
09:54:18 <tidoust> anssik: First face-to-face on Friday
09:54:53 <tidoust> ... 2 hour meeting. [going through agenda]
09:55:44 <tidoust> ... We'd like to start not with the proposed solution, but with use cases and requirements. Got some good contributions from Tomoyuki here for instance.
09:56:25 <tidoust> ... Then we'll go into more details on the proposal. If people have other proposals in mind, you're welcome to submit them for consideration.
09:56:42 <tidoust> ... I look forward to having a productive session. See you all on Friday!
09:57:11 <tidoust> RRSAgent, draft minutes v2
09:57:11 <RRSAgent> I have made the request to generate https://www.w3.org/2018/10/24-webmachinelearning-minutes.html tidoust
09:57:33 <anssik> RRSAgent, draft minutes v2
09:57:33 <RRSAgent> I have made the request to generate https://www.w3.org/2018/10/24-webmachinelearning-minutes.html anssik
09:57:53 <lin_> lin_ has joined #webmachinelearning
11:03:47 <dom> dom has joined #webmachinelearning
11:04:37 <dom> dom has left #webmachinelearning
11:06:56 <tomoyuki> tomoyuki has joined #webmachinelearning
11:07:20 <tidoust> tidoust has joined #webmachinelearning
11:09:12 <tidoust> tidoust has left #webmachinelearning
11:18:56 <sangchul> sangchul has joined #webmachinelearning
11:27:26 <zolkis> zolkis has joined #webmachinelearning
12:01:13 <sangchul> sangchul has joined #webmachinelearning
12:02:29 <Zakim> Zakim has left #webmachinelearning
12:04:27 <tomoyuki> tomoyuki has joined #webmachinelearning
12:07:09 <sangchul> sangchul has joined #webmachinelearning
12:07:41 <takio> takio has joined #webmachinelearning
12:13:45 <takio> takio has left #webmachinelearning
12:58:13 <sangchul> sangchul has joined #webmachinelearning
13:16:45 <sangchul> sangchul has joined #webmachinelearning
13:28:48 <tomoyuki> tomoyuki has joined #webmachinelearning
13:34:46 <sangchul> sangchul has joined #webmachinelearning
14:16:21 <sangchul> sangchul has joined #webmachinelearning
14:56:13 <zolkis> zolkis has joined #webmachinelearning
16:16:45 <zolkis> zolkis has joined #webmachinelearning
19:32:28 <zolkis> zolkis has joined #webmachinelearning