HN Explorer

31 comments

[ 0.20 ms ] story [ 90.6 ms ] thread
z3t4 

    let shape = [1,-1,1,1];
    return nn.add(
        nn.mul(
          nn.reshape(scale, shape),
          nn.div(
            nn.sub(input, reshape(mean, shape)),
            nn.sqrt(nn.add(nn.reshape(variance, shape), nn.constant(epsilon)))
            )
          ),
        nn.reshape(bias, shape)
      );
I find nested function calls like these very hard to comprehend. I would like intermediate variables and comments explaining what is happening and why. Make it flat.
sillysaurusx 

    let shape = [1,-1,1,1];
    # line up shapes.
    mean = nn.reshape(mean, shape)
    variance = nn.reshape(variance, shape)
    scale = nn.reshape(scale, shape)
    bias = nn.reshape(bias, shape)
    # calculate 1/sqrt(variance); add an epsilon to prevent dividing by 0
    epsilon = nn.constant(epsilon)
    inv_var = nn.rsqrt(nn.add(variance, epsilon))
    # batchnorm, I think?
    x = nn.sub(input, mean),
    x = nn.mul(x, inv_var)
    x = nn.mul(x, scale)
    x = nn.add(x, bias);
    return x
I don't know why they didn't use nn.rsqrt, since it's a common operation. In fact it's so common it's a builtin op: https://www.tensorflow.org/api_docs/python/tf/raw_ops/Rsqrt

I'm skeptical the reshaping is even necessary. Any reasonable framework should broadcast those shapes automatically.

baybal2
No, not yet another useless API. W3C is killing itself by allowing the extension of the web standard in a manner "Extend It Until It's Dead"

Microsoft should've not been allowed to join the W3C.

detaro
> the extension of the web standard

It's a community group document, so not a standard and not on the way to become one.

baybal2
But their goal is very much to include it into the core standard.

Second, W3C function as a standard promoter is compromised when it allows its members to make use of draft/private proposals as ersatz standards.

detaro
It's a common pattern. E.g. IETF has "informational" and "experimental" RFCs, which also are not standards, and I'm not aware of major problems with that.
throwaways885
Genuine question, but do you really believe this is the final form of the web? That we're done and it's perfect forever?
lwansbrough
iOS is still missing basic web app functionality. Can W3C focus on actually getting their proposals adopted instead of just churning out more crap that only has a future in Chrome?

I know this is mostly an Apple problem but I’m proposing the working group do more to apply pressure if it all possible.

pjmlp
Web in 10 years = ChromeOS
UnpossibleJim
Doubly so, now that schools run on it. That base familiarity carries further than people realize. You start pushing that OS in when kids are 5-10 and they'll just think that's what a computer should be. My kid would if I didn't "force" him to look at his library app on my linux box.
anaganisk
Yeah now let's bring user tracking and unnecessary GPU abuse, right into browser!
gridlockd
What "tracking" or "GPU abuse" do you expect this API to bring to the browser, that isn't already enabled by existing APIs?
anaganisk
The sad part is it will be used for more focused targeting, rather than for benefit, we have people abusing Canvas and audio context. I dont want companies to timestamp me by my face, to prove I like their content, or a dumb captcha API that decided im not a human because I'm darker. No please.
gridlockd
This API does not implicitly give access to the camera or photos. The things that you are concerned about are already possible by doing the NN part server-side, or less efficently in software. If anything, keeping the NN part on the client allows for more privacy, not less.
kzzzznot
Not saying I agree or disagree with gp’s sentiment, but perhaps section 2.1.1

> The application is watching whether she is in front of her PC by using object detection

gridlockd
You can already do that with Javascript, it is just less efficient than using the hardware meant to accelerate NNs.
AlexAltea
Why making this a Web API, rather than a library? We have WebAssembly threads which allows high-performance multi-threaded CPU code and WebGPU for general purpose GPU code.

Please correct me if I'm mistaken (not a ML expert): On top of these existing APIs, we can build efficient neural networks. Of course, we wouldn't be able to use specialized hardware such as Google's TPU or Graphcore's IPU, but I doubt there's a use-case for those in web.

Is there any non-GPU/CPU based consumer hardware aimed at fast neural networks that I'm missing?

EDIT: Seems "NPU"s have been slowly arriving for the past 2 years to iOS/Android-based devices... Maybe there's a raison d'être for the WebNN API after all.

pjmlp
Yes, stuff like https://developer.android.com/ndk/guides/neuralnetworks
AlexAltea
Thank you, I didn't know that! It feels scary to see that all these NN APIs are appearing around the fact that we cannot talk to DSPs, NPUs, GPUs, CPUs under a same language. By creating such highly-specialized APIs it feels like "giving up" on ever accomplishing that.

Decades ago, compiler developers solved dealing with multiple CPU architectures, by creating languages that succinctly expresses their common traits: Doing arithmetical-logical operations and accessing memory. Languages like C abstracted us from CPU registers, stacks. Optimization passes like vectorization abstracted us from different ISA extensions.

Now it feels it's the same issue all over again. Is it really hard to simply express: `x' = W*x + b` and let the compiler target the CPU/GPU/NPU as needed?

pjmlp
That is what stuff like SYSCL, CUDA and Halide allow for, but I guess the Web can have only one and WebAssembly won't be exposing what they need to work with anyway.
sillysaurusx
I kind of like the idea of having a common API for the common neural network ops, if only as an intermediate layer between pytorch and tensorflow.

But that's more an expression of API design than any particular reason.

Once TPUs can run JS (which may be coming sooner than it seems), it might also be reasonable to write a "webpage" which TPUs then "execute".

I think neural networks in general need to learn to be more flexible. Right now it feels like you're carving a network out of marble. It should be more like clay.

More practically, tensorflow.js does run in the browser; you could port this API to a tf.js backend.

tpetry
The needed models can be quite large. Do you really want to download a large model on every website making use of the new proposed apis?
AlexAltea
I don't think the WebNN API is trying to solve this problem, are they? Skimming through the specification, WebNN provides a way of describing the topology of a neural network, and compile+run it on different hardware (DSP, NPU, GPU, etc). However, the parameters of the network (which are by far the largest part) have still to be provided.
zerop
RNNNoise for Noise Suppression... RNNNoise is not having much activity since quite sometime.. I think the main guy joined Amazon.. not sure..
FridgeSeal
In today’s “new and unnecessary things nobody asked for” news segment.

Controversial, but stapling in unnecessary machine learning as a web api when we already have enough performance issues with web apps as it is. Anyways, why go to the effort of this, when if you really wanted to run ML models on the front-end you could use ONNX (a format designed already for portable network serialisation) and WASM.

ajtulloch
This sucks! Please don’t standardize this.