Google Is Partnering With Raspberry Pi To Create AI (zdnet.com)

Google is planning to bring artificial intelligence and machine learning tools to the diminutive Raspberry Pi this year. The Raspberry Pi Foundation said in a statement, "Google is going to arrive in style in 2017. The tech titan has exciting plans for the maker community." ZDNet reports: The advertising-to-cloud-computing giant intends to make a range of smart tools available this year, according to the Foundation. "Google's range of AI and machine learning technology could enable makers to build even more powerful projects," it said. Google has developed a huge range of tools for machine learning, IoT, wearables, robotics, and home automation, and it wants Raspberry Pi fans to fill out a survey that will help it to understand what tools to provide. The survey mentions face- and emotion-recognition and speech-to-text translation, as well as natural language processing and sentiment analysis. "The tech giant also provides powerful technology for navigation, bots, and predictive analytics. The survey will help them get a feel for the Raspberry Pi community, but it'll also help us get the kinds of services we need," said the Foundation.

Comments

Phone Home?

[iliketrash]

Of course, the Google-powered Raspberry Pi devices will all phone home. Right?

Re:Phone Home?

[swillden]

Of course, the Google-powered Raspberry Pi devices will all phone home. Right?

Not necessarily, it depends on what you use. Many of the AI tools are inherently cloud-based, relying on Google's models for image recognition, etc. Those are too big and too complex to run outside of a data center. But you can also build your own TensorFlow models and use Google's cloud service to train them (a computationally-expensive process, which Google has custom hardware to accelerate), then download the trained model and execute it on the RPi (or a different machine), offline.

Re:Phone Home?

[thegarbz]

Do you expect to run Google's AI software complete with data centers on your Pi?

This is Google giving access to its APIs there's no phoning home in the sense that you are being monitored.

AI?

[Camel Pilot]

Is it a bit presumptuous to call it AI? Machine learning yeah OK but AI? Really.

Sure

[OneHundredAndTen]

Like the other dozens of projects and partnerships that have promised pretty much the same over the last 40+ years.

Re:Sure

[Jason1729]

You're right about that over exaggeration except you're the one doing it. The Apple II came out 40 years ago this April. The Altair 8800 article that inspired Bill Gates was years old by that point. This is also the 40th anniversary of the TRS-80 and Commodore PET.

It would be a smaller exaggeration than your nugget to say there were 1000 different *models* of home computer on the market 40 years ago.

Simple: Cloud

[DrYak]

This is google.
Their AI offering don't have any installable component, not even on clusters.
They only provide APIs.

So the 1.2 GHz AArch64 running on your Pi3 will only take care to connect over HTTPS to some RESTful API.

And Google is trying to attract developers to their API as fast as possible, to avoid being beaten to the hackers IoT market dominance by other competitors.
(e.g.:
Microsoft has announced they open API access to their Bing/Cortana offerings,
Alexa from Amazon has even been featured on a few instructables.com on Raspberry Pis,
Houndify is also entering the market with an API specially oriented to easilly add Voice-to-Meaning to 3rd party apps, etc.)

Re:Simple: Cloud

[LifesABeach]

Took the survey, 1-am I a maker, 2-my age. Yup, that's all they are interested in? Hell of a survey.

Hay H1B Zombies, I know that putting AI into a rPi is not on your test, so go home and figure it out yourself. Impress us from Bombay.

Re:Simple: Cloud

[swillden]

This is google. Their AI offering don't have any installable component, not even on clusters.

One of their offerings is TensorFlow training. When it's done you download the model and run it locally.

API-only

[DrYak]

In fact, google doesn't even provide any installable AI component.
All their offerings are cloud API only.

So by definition, it only exclusively works by phoning home.

Voice-to-Meaning

[DrYak]

The "Voice-to-Meaning" capability of some AI clusters like Google's
(or Amazon's Alexa, soundhound's Houndify, Apple's Siri)
are quite impressive already.

The ability to have a (more or less) natural conversation with the assistant, and cluster running somewhere in the cloud being able to decode what you want it to do.
And such capabilites (thanks to a documented cloud API) can indeed be leveraged by any Raspberry Pi homebrew project, as long as an internet connection is available.

Re:Voice-to-Meaning

[religionofpeas]

So are you, except your database is smaller.

Better: NO cloud

[fyngyrz]

This is google.
Their AI offering don't have any installable component, not even on clusters.
They only provide APIs.

Keep your eye on the open efforts for TTS and STT. Those are where our best hopes lie.

Not in fixing ourselves to the rear bumper of some corporation with a bloody chain made of links of extracted personal information.

C'mon, people. If we're smart (I know, I know) we will wrest back control of our own destinies.

(Oh, and BTW, Google, that survey was incredibly lame. Who wrote that? A fourth grader?)

Re:Better: NO cloud

[LifesABeach]

Why not place the AI into a 128G thumb drive? When 256G thumb drives drop in price, move to that size?

Re:Better: NO cloud

[religionofpeas]

AI requires computation, not storage.

Re:Better: NO cloud

[LifesABeach]

I guess it boils down for A.I. students to publish which type and their respective sizes of problems can be handled by the Raspberry Pi? And as speed and memory size increase, then folks could have their own A.I. ?

Nope

[fyngyrz]

Keep your $50 dumbphone turned off or in Airplane Mode when you're not actively using it, and the ability to track you dwindles to essentially zero.

No. The only way to absolutely stop a tracking system from following you via your phone is to take the battery out. When it's "off", that doesn't mean it can't listen or use the radio. That "off" button is a softkey. Not a power switch. You think "I'm turning it off", it's going "hey, user pressed a button... now, what shall I do in response..."

However, no electrons available is a complete show-stopper.

Re:AI with only 1.2 GHz!!!

[AHuxley]

Broadband. The device is like a camera, microphone, sensor or input. The data is then sent to a supercomputer for the AI magic and profit.

Re:AI with only 1.2 GHz!!!

[allo]

Your mind does less operations per second.

TensorFlow Already Runs on a Pi Locally

[NotWallaceStevens]

TensorFlow Already Runs on a Pi Locally. There are plenty of projects out there working on this. I wonder what will be new here--better optimizations? I hope they address the use case of a Pi away from WiFi better.

Open Speech-To-Text

[DrYak]

Keep your eye on the open efforts for TTS and STT.

Speaking of which: I happen to have done a bit of online search about this subject (non-cloud STT).

Two things :

- Old style statistical based approach (N-grams and Hidden Markov Models) :
CMU Sphinx is an open-source implementation.
These older approaches are rather lighter on the processing power.
If you limit requirement (e.g.: only recognize a 1000-word vocabulary) it is possible to run it on embed CPUs.
There are people getting CMU Sphinx on Raspberry Pi3.
There is Open Ears which is a STT framework using Sphinx that is 100% running on iPhone and iPad.

- 2010s style AI (using Deep Neural Net, Recurrent Neural Nets like Long Short-Term Nets) - similar to what Google and co use on their networks :
Kaldi and EESEN based on it.
Of course these kind of approaches require way much more processing power than what available on most embed platforms.
But as they are opensource, you could install it on your server rack in your basement, and have the Raspberry Pis stream OPUS-compressed audio over Wifi of what they need processed.

Will it beat Google ?
Nope, due to less processing power and specially much smaller amount of training material.

Could it be used to add speach capabilities without needing to stream every single noise to Google and Bigbrother ?
Certainly.

Scale problems

[DrYak]

I guess it boils down for A.I. students to publish which type and their respective sizes of problems can be handled by the Raspberry Pi?

Most of the time : Nope.

It boils down to the fact that most complex A.I. problems are nowadays tackled using neural nets.
Which require usually orders of magnitude more power that what's available on a mere embed CPUs.
You usually want a couple of servers in your basement each equipped with graphics cards running the neural nets over some GPGPU solution (OpenCL, CUDA, etc.)

And that's without taking into account the massive datasets needed to train said neural nets.
e.g.: for Speach-to-Text
For Google, with thousands of users all streaming their voice commands to the mothership, they have a huge corpus to refine their models.
For us random geeks, we basically are limited to a few small open corpuses like voxforge.

So there are very few AI thinks that you can directly accomplish on the RaspPi itself.

Cheap

[DrYak]

If that's true then why is this even a Pi thing? If the heavy lifting is in the cloud then my phone or my watch can do it.

The RaspPi are cheap.

If you want to do tons of things (like use AI for voice commands in your home, image recognition on your security cams, etc...)
you either :
- buy an expensive phone or smartwatch for each.

or:
- buy a cheap RasbPi for each.

As the number of project increase the order of magnitude in price between a Pi and a smartphone start to really get noticeable.

Or is the problem that all developers are working on Pi hardware now? Where did all the devs go??

So yeah, most of the "hack at home" style of developers currently experiment with RaspPis.