Slashdot Mirror

You should see Tensorflow, a huge steaming pile of crud built on python and its legacy libraries.

Tensorflow has bindings for other languages. For instance: Tensorflow C++ API, with no Python needed.

TensorFlow?

There's no API. The models they used, data for training these models, and even pre-trained model weights are all open sourced. This particular implementation runs in client-side JS. A TensorFlow implementation in Python / Jupyter notebook is also available for local use.

Your assertion that Moore's Law is dead is akin to a similar assertion over 100 years ago that "everything that can be invented has been invented", and that the patent offices should be closed.

Go download Google's open source Tensorflow (an AI Machine Learning library), and try some real machine learning on real-time sensors and data streams. You'll quickly realize the highest end workstations can't keep up.

Now delve into a bit of devices physics. The easy gains in speed for silicon transistors have been made. There are still advances to be made, but different device physics that allow switching into the terahertz might just reset the clock on Moore's Law, which is just what's needed in all sorts of fields, such as AI.

Honestly, the only way that I see this happening is if Google decides to make their AI interface open source. Which they might do as a public service -- but we're still playing in Google's sandbox.

You mean like this?

I see Slashdot is too busy bitching about SJWs and global warming to read actual tech news anymore.

In answer to both Bruce Perens and destinyland, Google has open-sourced the TensorFlow library and created a public API to access their pre-trained instance of the library. Both of these announcements were made to a wider audience in March at Google's NEXT cloud conference, but it was publicly known since at least November 2015, when it appeared on Slashdot with a link to the source on GitHub.

That Slashdot posting got 37 comments. You people should be ashamed of yourselves.

Watch as AI replaces educated humans. It can. And it will. And sooner than you think.

Maybe, but the AI we have today doesn't seem all that intelligent. We're nowhere near the traditional idea of an AI as a self-aware consciousness.

For example, try going through the TensorFlow MNIST tutorials. Humans must supply significant input in order to get it to recognize Arabic numerals, something most of us do without any conscious thought.

Right, this isn't a general-purpose DSP but a custom ASIC designed to run their TensorFlow graphs efficiently.

But the fact that only companies and labs have access to this technology can represent a threat. First of all, we cannot know how money driven companies are going to use this novel technology. Second, this monopole slows down Progress and Technology.

The GNU project should do a bit more background research before starting new projects. Here are some links to open source deep learning tools. These are the same tools and libraries used by those "money driven companies" in their projects, including AlphaGo:

Caffe, widely used C++ deep learning framework.

Theano, widely used Python deep learning framework.

Torch, the software used by Google, AlphaGo and Facebook.

TensorFlow, Google's large scale machine learning framework.

CNTK, Microsoft's deep learning toolkit.

The idea that "the fact that only companies and labs have access to this technology can represent a threat" is patently absurd. Theano, Caffe and Torch are all open source and even Google has open-sourced its Tensor Flow platform which makes it easy to build new tools and run then, fast, on all the GPUs you can find. If you need to do this at scale and you're not the size of Google or IBM you can use Amazon's Machine Learning for AWS. There are many, many higher level toolkits out there that are available under licenses that are much less restrictive than GPLv3.

I've been out in industry for exactly a decade. I know who they are laying off. I would bet heavily that these are the guys that like doing things the way they have always done things and are content on not improving it. They're the drafters that refused to learn "that CAD thing". You see it all over Slashdot. You guys sure like things the way you used to do them. "Why kids these days don't need to learn Assembly".

I spent a recent layoff learning Python 3.4. It's near impossible to get people off of 2.7 at work or Matlab. Why? Because that's what they learned during undegrad and grad school and that's where everything is written. And they do have a small point, I'm don't have time to go back and re-write 50 years of working software. Once we as a society figured out Linear Algebra in Fortran we stopped messing with it. Numpy, Matlab, et al are just pretty BLAS wrappers.

However at impedes a lot of progress. At this point I feel like I'm in Office Space half the time:

1st Bob: What you do at Initech is you take the specifications from the software engineer and bring them down to the hardware engineers?
Tom: Yes, yes that's right.
2nd Bob: Well then I just have to ask why can't the software engineers use the hardware engineer's API?
Tom: Well, I'll tell you why... because... software engineers are not good at dealing with APIs...
1st Bob: So you physically take the flash files from the software engineer?
Tom: Well... No. The project lead does that... or they're e-mailed....

If you're doing things the same way you did them even a year ago, then some lazier person that does your job is currently writing a script to do it that way. So in 50 years we can all look back and laugh at "Those idiots used to do it by hand". If you write a script to save you 1 minute a day, that's 4 hours a business year. If you write a script to save you and all of your co-workers 1 minute a day. That's an additional 4 hours per head per year. Start adding that up over a decade or two.

It's entertaining to watch you guys not wanting to use new tools, I just started writing new tools to use the old tools I wrote. I could reduce my manager's headcount by 3-4 and keep the same work level output with an improvement in quality. Software engineers have already done that, it's what continuous integration is for. Then they got tired of dealing with merges, so they wrote tools to automatically do merges if everything tests out.

CGP Grey's "Humans Need Not Apply" is a good video on the current state of automation. While I don't share quite his outlook his statements about what is going on right now is dead on. (Humans' will just start building warp drives instead of dicking around with what we do now). If TensorFlow can pick those images out that accurately they sure as heck can read the graphs I used to have to read much, much better. Give me the picture of a tachometer trace and I could tell you what's wrong it your car. I don't need to hear it, see it or know what's going on.

Last night on SharkTank there was a guy that had a mobile app that could take your measurements 20% better than a professional tailor, just by taking some photos and doing some math. If you were hoping to be a tailor and spend time measuring people, I have bad news.

Engineers these days use Simulink. Finding Engineers that can Code is hard. So we taught the en

After looking at TensorFlow, I realized I'm not very prepared to use it, so I started taking Ander Ng's course and I'm in my 5th week now, and I feel like I'm getting a lot of it. I like the way he seems to have created this course to be fairly self-contained. Fors instance, although calculus shows frequently in the course, he is fairly open that he doesn't consider it to be a prerequisite and derived version of the equation is usually given whenever it comes up. Linear algebra is certainly required in the course, but the course provides nice refresher, and I actually learned it more firmly than I've gone through it previously in the past. (Maybe I'm more motivated than last time I went through it, though...)

Beowolf? Don't think so. If anything then maybe Spark with the map or reduce step being executed on the GPU, or better yet Tensor Flow. But as pointed out elsewhere, this chip is not even for that, especially next year when the new cards with NVlink blow away 980/Titan-X stuff of this year. No this thing is for drones, AR, or image recognition on embedded anything where power consumption and latency are the overwhelming factors. Otherwise graphics cards will outperform, or if latency is not a factor, then the whole thing can be offloaded to AWS or similar. Also, 16 bit (half-precision) floats normally bad for numerics are fine for neural networks with the bonus advantage of effectively doubling the memory bandwidth and problem size which are the current limitations.