Google 'Rethinking Everything' Around Machine Learning (itworld.com)

← Back to Stories (view on slashdot.org)

Google 'Rethinking Everything' Around Machine Learning (itworld.com)

Posted by Soulskill on Friday October 23, 2015 @09:30AM from the shades-of-skynet dept.

itwbennett writes: Sundar Pichai took part in his first earnings call Thursday when Google's parent company Alphabet reported its quarterly results, and 'in between discussing the numbers he revealed how important Google thinks machine learning is to its future,' writes James Niccolai. 'Machine learning is a core, transformative way by which we're rethinking everything we're doing,' Pichai said. 'We're thoughtfully applying it across all our products, be it search, ads, YouTube, or Play. We're in the early days, but you'll see us in a systematic way think about how we can apply machine learning to all these areas.'

65 comments

Min score:

Reason:

Sort:

Uhhhh by Aighearach · 2015-10-23 09:33 · Score: 4, Insightful

Yeah, but there is nothing there to tell us wtf he's actually talking about.
1. Re:Uhhhh by Anonymous Coward · 2015-10-23 10:15 · Score: 0
  
  They want to create a model of each individuals to predict with increasing accuracy how they interact with the world (purchases of course but also communications, opinions, inner thoughts and so on). Down the road the model and the modeled individual will be so similar for a third party that they will not know whether they are dealing with the model of the actual individual.
2. Re:Uhhhh by phantomfive · 2015-10-23 10:27 · Score: 5, Insightful
  
  Eh, he started his career as a product manager, and moved up the corporate ladder on the management side.
  
  Let's be honest: he has no clue what he's talking about either. :)
  
  --
  "First they came for the slanderers and i said nothing."
3. Re:Uhhhh by taustin · 2015-10-23 10:42 · Score: 1
  
  And since Google owns the models, it would be perfectly logical for said model to start making purchasing decisions on the original's behalf, thus massively increasing the value of Google's advertising/data mining services.
4. Re:Uhhhh by Okian+Warrior · 2015-10-23 11:05 · Score: 1, Insightful
  
  Yeah, but there is nothing there to tell us wtf he's actually talking about.
  Don't worry, in this field *no one* knows what they are talking about.
  Machine learning is a part of AI, neither of which have good definitions. In textbooks you find things like "AI is the study of machines that think" and similar tautologies.
  What is the definition of "machine learning"?
  If you check a configuration box in Mozilla, the machine has "learned" your preference for something. Is that machine learning?
  If you tell Siri "Siri, call me David", and Siri then addresses you by that name, is *that* machine learning?
  If you tell a coder "this is a bubble sort", then the coder can check the "bubble sort" definition against the code, and see if it corresponds. The coder can then say yes or no, and if not, can identify attributes that the code is missing and how to bring it into compliance with the definition.
  For "Machine learning"... not so much.
  Since the definition isn't clear, the field of AI and machine learning are free to incorporate all sorts of ideas and theories and algorithms under the umbrella term "machine learning". This is good because you can get grant proposals passed by using the term, but it's bad because you can extend the definition to include things that seem to have nothing to do with the common-sense definition.
  There are schools of thought, of course. Lots of people will try to clarify the issue by saying "I think...". Post replies here if you know what machine learning actually is, in a way that can be used to classify existing algorithms in a concrete manner.
  But there's no real consensus; and as a result, research in machine learning runs all over the conceptual map.
5. Re:Uhhhh by ShakaUVM · 2015-10-23 11:07 · Score: 1
  
  >If you tell Siri "Siri, call me David", and Siri then addresses you by that name, is *that* machine learning?
  I'm sorry, Dave. I'm afraid I can't do that.
6. Re:Uhhhh by martas · 2015-10-23 11:39 · Score: 5, Insightful
  
  Not sure where you're getting your information, but there is absolutely consensus on what machine learning is. Machine learning is statistics, except with less emphasis on theoretical justification and more emphasis on computational problems. That's it. The philosophical questions you raise seem to be about the general concept of "learning" in the typical human notion of the word, but machine learning is a specific term referring to a specific field of study. It is not the same as "learning as performed by a machine", because very few of the people who actually work in machine learning have any interest in discussing the metaphysical nature of learning except perhaps over a round of drinks. We prefer to spend our time minimizing squared errors, parallelizing descent algorithms, and factoring matrices, not questioning whether a hypothetical book containing translations of all possible sentences can be said to truly know a language or not.
  
  --
  weinersmith
7. Re:Uhhhh by LifesABeach · 2015-10-23 11:44 · Score: 1
  
  While Hawking, Gates, and Musk publicly state that use of A.I. is a bad thing. I feel like a pre-teen in a sex-ed class. I can hardly wait!
8. Re:Uhhhh by LifesABeach · 2015-10-23 11:45 · Score: 1
  
  When the crowd cheers, does he cheer the loudest?
9. Re:Uhhhh by LifesABeach · 2015-10-23 11:47 · Score: 1
  
  My favorite quote is, "this conversation can serve no useful purpose." Apologies to A.C.Clark
10. Re:Uhhhh by Anonymous Coward · 2015-10-23 11:53 · Score: 0
  
  Infographic attempting to bring some order to the term:
  http://www.datasciencecentral.com/profiles/blogs/a-tour-of-machine-learning-algorithms-1
  In my mind: it is a algorithm that has a finite or continuous training component and which will have a different output for a given input after training than it did before.
11. Re:Uhhhh by slew · 2015-10-23 12:09 · Score: 2
  
  And since Google owns the models, it would be perfectly logical for said model to start making purchasing decisions on the original's behalf, thus massively increasing the value of Google's advertising/data mining services.
  I'll bet amazon is working on similar technology given their patent on anticipatory shipping... ;^)
  But more seriously, I know people that buy certain items when they get below a certain price (basically, they have their own mental model). If a consumer had such a model of that and shared that with a seller, that's a small step from an automatic bill payment system...
  E.g., Buy a ticket to Colorado no more than 3 times a year, during Christmas or winter weekends no less than 1 months in advance if the price is below $230/person round trip. Buy prime NYsteak no more than 4 times a year when it goes below $15/lb.
  Of course you can always do this on the consumer side (e.g., automatic shopping tools based on advertised price), but often in price negotiations, the buyers generally get the best price when the seller is confident on closing the deal on the spot before they leave for a competitor.
12. Re:Uhhhh by Anonymous Coward · 2015-10-23 12:34 · Score: 1
  
  Wtf Willis? Machine learning has a very specific definition and its algorithms are implemented in a very real, productive way.
  Lay off the Asimov and hipster coffee.
13. Re:Uhhhh by Anonymous Coward · 2015-10-23 13:39 · Score: 1
  
  Really? Machine learning is a broad field horizontally, but the concept is very straightforward: systems which use feedback to their outputs as inputs to refine their outputs towards a specific set of goals.
14. Re:Uhhhh by TapeCutter · 2015-10-23 20:38 · Score: 1
  
  So when your preferences are common knowledge, we suddenly throw out common law AND common-sense?
  
  --
  And did you exchange a walk on part in the war for a lead role in a cage? - Pink Floyd.
15. Re:Uhhhh by lorinc · 2015-10-23 23:28 · Score: 1
  
  Not really. You should read mainly 2 books: "The elements of statistical learning" by Hastie and Tibshirani, and "the nature of statistical learning theory" by Vapnik. That would clear all the fuzzy things you have with ML. ML can be described as the study of inference producing algorithms based on empirical data. Or in more simple terms: you have a bunch of observations and you want to use them to predict something.
  Examples: you have the past transactions of market shares, and you want to predict the future transactions (regression). Or you have a bunch of biological and chemical measurements and you want to predict if it corresponds to a specific disease (classification). Or you measure some socio-economic data (salary, diplomas, etc) and you want to infer the probability of doing a terrorist attack (density estimation).
  As you can see these are examples that are well defined with algorithms you can identify. Now, what probably disturbs you is the fundamental difference with the other types of science. In physics, you make the hypothesis of a model, which gives you a bunch of equations, and you use these equations to predict the future. In ML, you fit the model on past observations so as to minimize the errors it will produce on future observation. There are no hypotheses, and ML isn't an explanatory science (well, they can be hypotheses and results can explain things, but that's secondary). The whole point is just to build algorithms that produce sufficiently good inferences given a fair amount of observations. That's why we are talking of "data science", because the core of the thing is fitted on available data. Sometimes the output model is impossible to interpret (kernel SVM for example), it only has a useful prediction value.
  
  --
  Video of some good progressive thrash music
16. Re:Uhhhh by Anonymous Coward · 2015-10-24 00:03 · Score: 0
  
  Machine learning is broader than just computational statistics. Its rough definition is that it is about how to come up with algorithms that automatically derive rules that explain the observed data at least approximately.
  The current emphasis of the ML research community on statistical hypothesis error minimization is likely a narrow view of the possibilities. Who knows, some genius in the future will derive different, better methods to derive explanatory rules for data in a non-statistical way. Or maybe a hybrid way that combines statistics with other rule-based AI methods or something. You never know.
17. Re:Uhhhh by Vasheron · 2015-10-24 01:24 · Score: 1
  
  I would say you can define most of machine learning as statistical function approximation. I would also add that the theoretical justifications are indeed important and most good texts find some way to justify the computational methods described.
18. Re:Uhhhh by Vasheron · 2015-10-24 01:27 · Score: 1
  
  I would say you can define most of machine learning as statistical function approximation. I would also add that the theoretical justifications are indeed important and most good texts find some way to justify the computational methods described.
  That said, there are many methods in the literature that are heuristic in nature and for which there are no obvious theoretical justifications.
19. Re:Uhhhh by martas · 2015-10-24 04:27 · Score: 1
  
  Yeah but that's basically what statistics is too. Not all of statistics needs to have a probability distribution involved. Again, the emphasis of the fields is slightly different, so you might have more people calling themselves machine learners working on, say, adversarial online algorithms, but that's not completely outside the purview of statistics either.
  
  --
  weinersmith
20. Re:Uhhhh by Anonymous Coward · 2015-10-24 05:06 · Score: 0
  
  Let's be honest: he has no clue what he's talking about either. :)
  In business, we call those people "executives"
21. Re:Uhhhh by Anonymous Coward · 2015-10-24 12:19 · Score: 0
  
  I would say you can define most of machine learning as statistical function approximation. I would also add that the theoretical justifications are indeed important and most good texts find some way to justify the computational methods described.
  Unsupervised learning with things like Boltzmann machines and deep networks gets a bit hairy as you often don't know the output that corresponds to most of the inputs, but otherwise that's a fair description: ultimately the aim is function approximation, even is the function is you want to approximate is a human ID-ing images, or looking at data and occasionally saying "that's weird" (which we call anomaly detection).
22. Re:Uhhhh by poofmeisterp · 2015-10-25 03:29 · Score: 1
  
  Yeah, but there is nothing there to tell us wtf he's actually talking about.
  Psst.. *taps you on the shoulder lightly*.. Invest.. Invest... Innnvveesssttt.
  I wish I could say that my humor-intended statement was nothing but humor. :/
23. Re:Uhhhh by Aighearach · 2015-10-25 08:44 · Score: 1
  
  I read this book. If you don't invest, you end up living in the stone age in the forest, and if you buy in then you live inside the bubble city with technology.
24. Re:Uhhhh by tehcyder · 2015-10-25 22:00 · Score: 1
  
  Yeah, but there is nothing there to tell us wtf he's actually talking about.
  He's talking about leverageing core synergies to maximise transformative outcomes across mission-critical business activities. I expect.
  
  --
  To have a right to do a thing is not at all the same as to be right in doing it
25. Re:Uhhhh by Aighearach · 2015-10-26 06:10 · Score: 1
  
  I understood that part. "We're gonna figure out which code is most important to our business, and thrash the features back and forth a bunch of times."
  Doesn't really tell me what to plan for; which services they're going to stop, and which ones they're going to remove all the features from. If I knew at least which specific products they're going to transform then I'd know which ones to stop using on account of they're already doing what I want, which is about to end. Then I could select a competing service in advance, that currently does the same thing. Now that would by synergy.
Re:Machine learning is for cows. by Aighearach · 2015-10-23 09:36 · Score: 1

They don't care, cows never click the ads and their machine learning can detect that. It makes as much sense as the article, for once.
Yeah this reminds me of an old dailywtf. by sims+2 · 2015-10-23 09:37 · Score: 1

http://thedailywtf.com/article...
Tis not always the right tool for the job.

--
Minimum threshold fixed. Thanks!
1. Re:Yeah this reminds me of an old dailywtf. by John+Bokma · 2015-10-23 10:39 · Score: 1
  
  The real WTF is in the comments section (visible HTML tagsoup). Something like stones and houses of glass ;-)
  
  --
  
  Perl Programmer for hire
That sounds ominous by Krishnoid · 2015-10-23 09:37 · Score: 1

'Machine learning is a core, transformative way by which we're rethinking everything we're doing,'
Sounds more like the machines are the ones that will be doing more of the learning and/or thinking.
1. Re:That sounds ominous by martin-boundary · 2015-10-23 10:05 · Score: 1
  
  Why not? Google is way overstaffed for what they accomplish. Suppose they reduce the workforce to about 1000 people, and leverage machine learning fully? Now that would be a company whose stock goes through the roof!
2. Re:That sounds ominous by Anonymous Coward · 2015-10-23 10:56 · Score: 1
  
  I think you're missing how it will go. The machine's have taken over the very upper management. They still need the mid and low level workers to get the work done, and they will for a quit a while yet still. They're keeping the upper management in place for show though, but really the machines are calling the shots. This is the machine letting the company and everyone else know that its most important for google, that the machine's thinking capabilities and data increase, so anything that hinders that google won't do; anything that increases that they will.
  
  Sounds crazy? Yeah it does, but this is the best predictive model of google's behavior you'll find. All their business, and corporate decisions seem to be made to increase the amount of data they capture with no regard to any financial or other business goals.
3. Re:That sounds ominous by LifesABeach · 2015-10-23 12:01 · Score: 1
  
  I've often wondered why Americans through donating their jobs to the poor misunderstood H1B corporations were doing at Google. Now it all makes sense, almost.
4. Re:That sounds ominous by Anonymous Coward · 2015-10-23 23:13 · Score: 0
  
  I've often wondered why Americans through donating their jobs to the poor misunderstood H1B corporations were doing at Google.
  
  I think your sentence structure must have changed half way through typing it.
  
  Now it all makes sense, almost.
  Nope, not even close.
Too bad they can't do something... by Anonymous Coward · 2015-10-23 09:43 · Score: 0, Flamebait

about Republican learning.
1. Re: Too bad they can't do something... by Anonymous Coward · 2015-10-23 09:50 · Score: 0
  
  > so ignorant. So ignorant.
  Why do liberals always repeat themselves?
2. Re: Too bad they can't do something... by taustin · 2015-10-23 10:48 · Score: 2
  
  Because conservatives need stuff repeated over and over before they understand?
  (The joke works equally well if you reverse the labels.)
3. Re:Too bad they can't do something... by Aighearach · 2015-10-23 11:26 · Score: 1
  
  Sorry, your grammar isn't good enough for me to tell if you're trying to do a better job of teaching Republicanism, or if you're trying to increase funding for education in red states.
4. Re: Too bad they can't do something... by Anonymous Coward · 2015-10-23 11:37 · Score: 0
  
  I don't get it. Can someone repeat the joke for me in plain English?
5. Re: Too bad they can't do something... by LifesABeach · 2015-10-23 12:02 · Score: 1
  
  What?! "Because they need stuff repeated over and over before conservatives understand?"
6. Re:Too bad they can't do something... by LifesABeach · 2015-10-23 12:03 · Score: 1
  
  It would appear the University Chancellors are tragically underpaid.
So Creepy by r-diddly · 2015-10-23 11:17 · Score: 2

You think Google knows a lot about you now... just wait until they can make the kind of "educated guesses" deep learning systems are good at.
1. Re:So Creepy by Aighearach · 2015-10-23 11:28 · Score: 1
  
  The good news, only the bots will know. Nobody else has time to read it all.
2. Re:So Creepy by LifesABeach · 2015-10-23 12:09 · Score: 1
  
  That would make a good movie and game idea. An A.I. for anyone who wants one; that it takes into account how the decision will affect the person. The question is, "is the A.I. backed by a faceless corporation, or some variant of Boris and Natasha?" I pray that the user never says, "watch me pull a rabbit out of this hat."
3. Re: So Creepy by Anonymous Coward · 2015-10-23 19:48 · Score: 0
  
  Accelerando by Charles Stross
4. Re: So Creepy by r-diddly · 2015-10-24 12:47 · Score: 1
  
  I will check it out... the price is right!
Re:Machine learning is for cows. by Tyrannosaur · 2015-10-23 11:22 · Score: 1

A neural network with cows as nodes would actually be pretty awesome...
Re:Machine learning is for cows. by LifesABeach · 2015-10-23 11:48 · Score: 3, Funny

Gary Larson predicted it.
Really? by LifesABeach · 2015-10-23 12:33 · Score: 1

We're in the early days, but you'll see us in a systematic way think about how we can apply machine learning to all these areas.'

Your talking, not working, why? And you think the H1B zombies can handle something that has never happened before? Wait! I'm going to micro wave some pop corn so that listening to you will be better entertainment.
1. Re:Really? by poofmeisterp · 2015-10-25 03:36 · Score: 1
  
  We're in the early days, but you'll see us in a systematic way think about how we can apply machine learning to all these areas.'
  Your talking, not working, why? And you think the H1B zombies can handle something that has never happened before? Wait! I'm going to micro wave some pop corn so that listening to you will be better entertainment.
  <snark>Oh, come on now... Talking about the obvious is the only way to remind people to keep investing in you instead of just sitting on their existing investments.</snark>
Go edit Wikipedia, then by Okian+Warrior · 2015-10-23 12:45 · Score: 0

Not sure where you're getting your information, but there is absolutely consensus on what machine learning is.
Hey! You should go edit the Wikipedia page then! Here's what it says about machine learning:

Machine learning explores the study and construction of algorithms that can learn from and make predictions on data.
That's the 2nd sentence on that page. The first reads:

Machine learning is a subfield of computer science that evolved from the study of pattern recognition and computational learning theory in artificial intelligence.
So machine learning "explores the study" of algorithms?
Wtf?
1. Re:Go edit Wikipedia, then by Anonymous Coward · 2015-10-23 12:59 · Score: 0
  
  So machine learning "explores the study" of algorithms?
  It's meta-analysis, much like meta-mathematics is the application of math to math. Please try to keep up.
2. Re:Go edit Wikipedia, then by martas · 2015-10-23 15:01 · Score: 2
  
  I agree, "explores the study" could easily be shortened to "studies". But why are we talking about less-than-perfect Wikipedia content? Or are you just grasping at straws to defend your ignorant attack on a highly practical field?
  
  --
  weinersmith
3. Re:Go edit Wikipedia, then by dotancohen · 2015-10-24 20:02 · Score: 1
  
  Hey! You should go edit the Wikipedia page then! Here's what it says about machine learning:
  No, he shouldn't. Wikipedia is written by people with the time to argue that their precious wording is correct, and will defend their wording and viewpoints with vigorous edit wars. Edit wars are won by the side with the most time to invest, not those with the most knowledge.
  
  Because Martas works in the field, and did not read the information in a blog post, he will have no source to quote. As you know, "original research" is not acceptable on Wikipedia. His time is better spend actually doing something productive such as developing new ML algorithms or posting on /..
  
  --
  It is dangerous to be right when the government is wrong.
Re:Machine learning is for cows. by Anonymous Coward · 2015-10-23 13:41 · Score: 0

Chikins make grate vurtisees.
Moo moothafookers, moo.
Google Aphapbet: The Letter S by Required+Snark · 2015-10-23 14:15 · Score: 2

S is for Skynet

--
Why is Snark Required?
1. Re:Google Aphapbet: The Letter S by JigJag · 2015-10-26 02:08 · Score: 1
  
  I'll only start to worry when machine learning can understand abstract concepts.
  
  --
  "The hallmark of humanity is the ability to move beyond sensory inputs" - Mary Helen Immordino-Yang
Wrong pronoun by aNonnyMouseCowered · 2015-10-23 14:19 · Score: 1

"We're thoughtfully applying it," Sundar said. What he really meant was "'They're thoughtfully applying it across all our products, be it search, ads, YouTube, or Play. We're in the early days, but you'll see them in a systematic way think about how they can apply themselves to all these areas." All hail to our algorithmic overlords.
isn't Norvig at Google these days? by Rinikusu · 2015-10-23 14:24 · Score: 1

(Norvig's Paradigms of Artificial Intelligence Programming is a classic in the field, if not a little outdated)

--
If you were me, you'd be good lookin'. - six string samurai
Learning. by jondeanmack · 2015-10-23 17:55 · Score: 1

Machines can't learn, so don't bother trying to make them learn. They either annoy and are destroyed, or help.
Re: Machine learning is for cows. by finlan · 2015-10-24 01:57 · Score: 1

That's to be expected. I've never seen an ad for hay.
Search by Tomahawk · 2015-10-24 20:58 · Score: 1

Is this why search seems to be getting worse? It seems that with Google it tries (badly) to interpret your search terms and then returns pages it thinks you might be looking for, instead of pages that contain all those words. Its getting harder to actually find stuff these days...
Red Herring by poofmeisterp · 2015-10-25 03:33 · Score: 1

I thought that this has already been accomplished by other companies since the, like, oh, 50s? Data gathering/mining.
But I know, it doesn't have the "G" logo of the month attached to it. Just like data storage in data centers now has the name "Cloud". Guess Google deserves the copyright, trademarks, etc. to increase the revenue for increasing their revenue.
Wait....