Google's AI 'TensorFlow' Software Is Coming To iOS (cnet.com)

An anonymous reader writes: Google published an early version of TensorFlow that adds support for iOS. TensorFlow is "neural network" software that lets computers process data in a way similar to our own brain cells. Google CEO Sundar Pichai recently said it advances machine learning capability by a factor of three generations. With the software running on your iPhone, its capabilities will allow for more sophisticated apps to run on iOS. We can expect the apps to be released later this year and into next year from Google and others who use TensorFlow. Some of the tasks TensorFlow can allow for include being able to recognize subjects in photographs or being able to teach your phone what a particular object looks like, which is what another neural network software project called MemKite aims to do. Google has released its TensorFlow software as open source, where anyone can use or modify it for free.

Comments

And I thought

[johnsmithperson123]

That iPhones were immune from the Robot Apocalypse because it wasn't on the App Store. I was wrong.

WOW A COMPUTER BRAIN?

[Anonymous Coward]

what will they think of next?!?

So many words meaning nothing

[Anonymous Coward]

Wow, that summary is terrible, I'm not going to bother reading the article because the summary is so bad. It advances machine learning by 3 generations? Compared to what? It already exists so isn't it current generation and the previous one 1 generation prior? You can't really skip generations. Also, it processes data the same way our brains do? I was unaware we understood how our brains processed data. Somebody should inform the neuroscientists, they'll be excited to find this out. Granted contrary to their name, those of us actually versed in machine learning know that neural nets have very little to do with how brains function.

Now get off my lawn.

Re: So many words meaning nothing

[hackwrench]

I agree with much of what you say, but we have a pretty good grasp of how the brain processes data. They tend to get a little confused on the terminology, like for some reason only the electrical firing qualifies as "signaling" and not apparently chemical transferrals. They have any number of ways of getting data to replacement appendages and even wire up primitive artificial eyes. They have implants for some seizure conditions. There's no firm lines between the knowns and the unknowns, though.

Re:So many words meaning nothing

[ShanghaiBill]

It advances machine learning by 3 generations? Compared to what?

They mean three Moore's law doublings. 2^3 = 8, so it is roughly 8 times as powerful as the previous state of the art, in terms of computation per watt. But that is only if you use their ASIC accelerator, which the iPhone doesn't have, so the "three generations" doesn't really apply here.

Also, it processes data the same way our brains do?

Basically, yes, artificial neural networks (ANN) work on the same principles as a brain, in the same way that 747s work on the same principles as hummingbirds.

I was unaware we understood how our brains processed data.

We don't understand how brains work at a gestalt level, but we have a good idea of how neurons and networks of neurons work. ANNs are modeled on biological neural networks (BNNs), can do similar (albeit simpler) things, and even make similar errors.

Somebody should inform the neuroscientists, they'll be excited to find this out.

Most neuroscientists are already well aware of research on artificial models of neurons.

Re:So many words meaning nothing

[Anonymous Coward]

Moore's law says nothing about computational power.

Re:So many words meaning nothing

[Anonymous Coward]

There is no 3 generations of machine learning; the summary is horseshit. On their blog Google claimed their TensorFlow accelerator called a TPU (which is rumored to only accelerate their inference library) "delivered an order of magnitude better-optimized performance per watt for machine learning", which they equated to 3 generations of Moore's law (what!?).

We don't know what benchmarks they ran or what they were comparing their results against. Did they do TPU accelerated loads vs loads running only on the host CPU? Did they compare against GPU or FPGA acceleration? Did they compare TPU acceleration against other machine learning libraries? Do those other libraries also have hardware acceleration that was benchmarked?

The TensorFlow news has been a shitty game of telephone by people who spew industry jargon but don't know what it actually means. I should go design a gzip ASIC and claim I've advanced compression by x-many generations.

Re:So many words meaning nothing

[Anonymous Coward]

They meant Koomey's law. But since Google isn't in the semi industry they decided to use layperson jargon.

Re:So many words meaning nothing

[TheRaven64]

I was unaware we understood how our brains processed data

We understand how neurones work, which is roughly the equivalent to understanding how transistors work and saying that you understand a CPU. We also know that you can approximate any mathematical function with a network of neurones. The accuracy of the approximation depends on the complexity of the network and with modern computers you can simulate very complex networks (not in comparison to an actual brain, but in comparison to the networks built by the people who worked on this stuff 50 years ago). The two big advantages of neural networks over conventional programs are:

• If you have a load of input and output from a function, you can train a network to approximate the function even if you don't understand what the function actually is.
• It's easy to trade accuracy for performance / power consumption in a neural network. It's hard with most algorithms to take an accurate algorithm and replace it with one that's 70% accurate but uses far less power. For a lot of pattern-matching things, a fairly high false positive or false negative rate is still useful.

There is a huge amount of hype around neural nets at the moment, because an entirely new generation of young computer scientists has learned that they can approximate any function (most of the old ones that knew this have retired or died).

Re:So many words meaning nothing

[110010001000]

Uh no. We don't understand how neurones work. You can have as many simulated "neurones" as you want but it will not make an artificial brain.

Spyware

[khchung]

Why would I want to install Google spyware on my Google-free iPhone?

Re:Spyware

[Jeremi]

Why would I want to install Google spyware on my Google-free iPhone?

To give the Apple spyware some company?

Re:Spyware

[thoromyr]

not that TensorFlow is necessarily spyware (I have no idea), but what "Apple spyware" are you referring to?

Re: Spyware

[Anonymous Coward]

I think the point was just that and you missed it. There is as much proof in the accusation that this is Google spyware as any Apple software is spyware.

Any real examples?

[Anonymous Coward]

Whenever machine learning comes up it's almost always identifying 2D photographs.. yeah, that's great, but it's not exactly my first thoughts when it comes to "machine learning"

Re:Any real examples?

[ceoyoyo]

Ever used Google translate?

Re: Any real examples?

[Anonymous Coward]

yeah, it's comical at Best...and down right horribley wrong with any difficult translations. Just try translating Mandarin, and you'll think it's just making random guesses.

Re: Any real examples?

[ceoyoyo]

It does a lot better than me at about 90% of the languages it knows. Maybe 95%. And it's faster than anybody.

Also, AI != better than human intelligence.

hype from google

[sittingnut]

other comments have pointed out the terrible confused inaccurate content of the summery here. and in article, and the claims google is making about this.

there is no new technological breakthrough here, that " advances machine learning capability by a factor of three generations", instead this is another implementation of a, long known, trade off that increase computation power/capability by decreasing precision.

this is hyped as how to "process data in a way similar to our own brain cells"(while real brain cells continue to be a mystery btw) or "AI" (whatever that vague, and useless, term mean or does nor mean), etc etc.

fact that google, from very top down, felt the need to hype such things, in totally exaggerated and inappropriate way, does tell something; google is in trouble new technology wise.

Re:hype from google

[Anonymous Coward]

there is no new technological breakthrough here, that " advances machine learning capability by a factor of three generations", instead this is another implementation of a, long known, trade off that increase computation power/capability by decreasing precision.

Actually, it's worse than that, and it looks like a submitter and/or editor error. That quote about 3 generations (and the associated /. link) refers to a hardware coprocessor designed by goog that is, from their declarations, what they run their in-house branch of the tensorflow software on.

The submission is thin on fact and heavy on the word salad, since 'hey, we tagged v0.9 RC0 with iOS support on github' would have been too terse, I suppose. And python3.5 support, OSX gpu processing support, nevermind more specific imporvements, are apparently too obscure for the submitter to mention. Because iOS, perhaps.

Re:hype from google

[SpinyNorman]

Yep, the summary is cringe-worthy. Tensor flow is just a framework that lets you easily build multi-step pipelines for processing multi-dimensional matrices (aka tensors). The matrices/tensors flow thru the pipeline, hence the name. The main targeted application is deep neural nets, and there are layers of functionality built into TF for building deep neural nets. There are a number of other preexisting open source frameworks that provide similar functionality. TF appears well designed (very modular, good for research), but it's no game changer.

Re:hype from google

[Anonymous Coward]

The whole idea that artificial neural networks process data like the brain isn't correct. At best, it's a simplification.

'Inspired by' might be a better term. And, like movies, 'inspired by' means 'nothing like'.

TensorFlow requires Python; iOS forbids Python

[thisisauniqueid]

A lot of the TensorFlow functionality is written in Python right now, for whatever bizarre reason (many TensorFlow pipelines take a huge performance hit by dropping out to Python after every training batch, in order to feed the next batch from a Python data structure). The TensorFlow team eventually plans to push more of the Python functionality down into C++-land, so you can build bindings for other languages (e.g. Swift/Obj-C), but this isn't currently possible. Since Python doesn't run on iOS (modulo a few hacky solutions like tinypy), I'd say the TF team has a lot of work to do before TF on iOS is a possibility. (Note -- this is for building arbitrary TF models in iOS, which would currently require Python, as opposed to executing runs through already-trained models, which can be done more simply by means of a TF graph serializer and de-serializer, which doesn't necessarily require Python, and already exists in some form for Android and other runtimes.)

Re: TensorFlow requires Python; iOS forbids Python

[Anonymous Coward]

so why would you even run the trainer on ios?

Re:TensorFlow requires Python; iOS forbids Python

[ceoyoyo]

If using Python between batches is a problem for you, make your batches bigger. Or use a C optimizer. Or use Cython for the training loop.

Re:TensorFlow requires Python; iOS forbids Python

[khallow]

many TensorFlow pipelines take a huge performance hit by dropping out to Python after every training batch, in order to feed the next batch from a Python data structure

Since most of the time is spent in the training batch not in the pipeline, so what? And let us note that the primary reason the functionality is in Python rather than C++ is because it is faster to develop in Python. CPU cycles aren't the scarce resource here.

Since hen does iOS forbid Python?

[SuperKendall]

iOS has had python apps for quite some time, a few years ago Apple decided it was OK to let things like that on the store.

You can easily ship an iOS app with a python library...

TensorFlow DOES NOT require python

[melted]

TensorFlow has a C++ API as well.

Re:TensorFlow DOES NOT require python

[thisisauniqueid]

Yes, of course it does, the C++ API is the one that Python uses. But there is a lot of functionality in Python that is not available in the C++ API, because a lot of rich features were quickly added in Python, rather than taking the time to develop them in C++: the Python API is not just a thin layer over the C++ API.

Does this mean

[LeadSongDog]

that OCR will finally work?

Plant scanning

[Anonymous Coward]

We need a mashup between every known plant maybe trained by wikipedia species. Then I can look at a plant and identify it on my phone. Reverse image search isn't smart enough for this because every leaf is slightly different but mostly green. AI could do this and tell hiking guides where harmful/notable plants are in real time.

Re:Plant scanning

[Incadenza]

You mean, like leafsnap?