This headline doesn't include the most important information. Russ was part of the big bang in deep learning about 10 years ago when he was Geoff Hinton's student. His name is on the 2006 papers about RBMs.[0] CMU is a great school, but Russ is not a random guy from CMU. He's a deep learning veteran who's done a lot of important work, which makes this a huge hire for Apple. They were lacking a marquee name for their AI team, and now they have it, and that's probably going to accelerate their recruiting and product development.
A marque name may be useful for Wall Street or the media but it seems unlikely it will help them recruit, for two reasons.
1) Machine Learning is one of the most quickly advancing fields of learning today, maybe ever. Having discovered a lot ten years ago likely isn't as important as what you have done now. The importance of the machine learning people that do AMAs on reddit (Lacune etc) is that they lead teams of people now who are working with streams of big data now, publishing state of the art results now. Which brings the second point...
2) Machine learning has advanced to its situation by being an incredibly open field, with arxive being the default repository for publications. Just doing this hire doesn't give any indication that Apple has relaxed its posture of restricting all information coming out of its research. This closed-to-publication position makes them much more unattractive to cutting-edge machine learning researchers than the presence of a single big name researcher could make them attractive.
(and I'm no cutting edge researcher myself, just an observer but I think you can find similar comments from real researchers in places)
This is fair, but as we've seen with their recent approach to Swift, Apple's shown it's at least possible for them to pursue a more open approach.
I'd like to think Maps and Siri have been a huge wake-up call to them that you can't apply the same closed-door approach to product development when your product require mountains of data to provide a good user experience. Then again, time will tell.
I wonder if they are thing the same thing along the lines of the Stratechery article pointing out the differing cultures required for services vs. devices. It may signal some sort of pivot, at least for their icloud based things, especially w/ the recent news of a reorg putting their services people all under a single roof/organization...
While I think you make valid points in general, they don't apply here for two reasons:
1) Russ has continued to do relevant and interesting work in the decade since those papers with Hinton. Here's his arXiv feed. The latest was in June 2016 I think.
2) Russ will keep his position at CMU, which means that he will continue to engage in interesting work that can be shared. I'm not sure how Apple's position on open-ness is evolving, and I'm not incredibly optimistic that they'll let people share their work, but Russ will keep one foot in the open community.
I'd never used Siri in the 5+ years I've been using an iPhone. Then I switched to Android for a month while awaiting the iPhone 7 (my previous phone screen cracked and could not be repaired a month before the new release).
After being impressed with "OK Google" I decided to try out Siri on my iPhone 7. I have been shocked how much better "OK Google" works than Siri. I think Apple has a fair bit of catching up to do in this area.
With OK Google you can basically just tell it what app to open, what to do, etc., and it works. With Siri, something as simple as asking her to play a song doesn't actually start playing the song, it forces you to interact with a clunky UI on the siri screen first, then it doesn't even open the correct song in Apple Music.
I'm not sure what "clunky UI" you need to interact with, you can play music without touching your phone. You literally just say "Hey Siri, play Two Weeks by FKA Twigs" and it will begin playing. It shows your request being transcribed when its made but that screen goes away on its own shortly after.
Most people's complaints about Siri as far as I can tell (via Twitter mostly, so it's a tech crowd) is that it's basically not Google Now. That is, Apple doesn't have a search engine so it can't answer general interest/trivia questions unless it's specifically related to a domain Apple has included (such as Wolfram Alpha or MLB for baseball).
Now many people think it should be able to do that, but that's certainly not what it was designed for. Siri is a "personal assistant" to help you do things on your device via voice. They've recently begun to expand these integrations by creating a Siri API which will probably expand to cover more domains over time.
>Now many people think it should be able to do that, but
>that's certainly not what it was designed for.
That's just a cop out though to say it isn't designed for that. Siri's biggest competitor is designed for it and does it very well on top of doing everything Siri does besides that (except perhaps wolfram and witty replies). Apple's hand is forced on the matter, and frankly their hands are tied without having an extremely powerful search engine under their belt. I doubt they are gonna get too far scraping Bing before MS steps in, if that's the route they decide to go.
It certainly doesn't do everything Siri does on iOS devices, because it can't. Apple's hand won't be forced because no one buys an Android device over an iPhone/iPad because Siri doesn't do the same things Google Now does. Google believes that is a feature that can sell phones, but I predict that the Pixel will never pose a threat to Samsung or Apple and likely won't sell much more than Nexus devices have.
Apple also has a good working relationship with Microsoft so if they wanted this functionality it wouldn't be via "scraping".
Oh well sure it can't on iOS devices, but that is Apple's doing, not Google's.
It does have full functionality on all GMS (Google Mobile Services) Android devices though (Sammy, HTC, LG, 1+, Sony, etc.). As for people not buying a device for it's assistant, I think that is an extremely short sighted comment. At least in the US where everyone has an iPhone and considers Siri near worthless, not many people use their assistant. And many of those that do are locked into a mindset that it can only do basic things/answer simple questions (i.e. remind me of x, whats the weather, whats score of the game, text nick hey). My friends head nearly fell off the other day when I asked my phone why people put nitrogen in their tires and it read back a perfect response without me even having to touch it.
As people break out of the "assistants are shitty" mindset in coming years, and especially as talking to your phone becomes more socially acceptable, I think there will be much more focus on who does it best. Without a doubt the future is going to be riddled with AI. In that sense to write off personal assistants as marketing gimmicks sounds as ludicrous as writing off the electric light bulb as a neat trick in 1850.
You missed my point, which was that Google Now can't do the things Siri can on iOS. This is important because Apple has aggregated a large and highly valuable user base. Contrary to your claim otherwise, this user base makes 2 billion Siri requests per week [1]. The fact that Google Now can't insert itself into this position works to Apple's advantage because that dataset will enable Apple to improve the product over time. It also means that everyone who wants to reach that audience will partner with Apple and/or use the newly available Siri API (as many have done), which will further add new functionality.
And if Apple really wanted to replicate the "answer general interest questions" use case with Siri it wouldn't require Apple to create a search engine, just a bizdev deal with Microsoft and some engineering work (probably from both parties).
I also never made any claims that assistants are shitty or would stay shitty. I said that Assistant/Google Now isn't enough differentiation to convince a customer to purchase a Pixel over an iPhone (or a Samsung Galaxy for that matter). The iPhone's competitive advantage doesn't rest on any one single feature or service. It isn't enough for Google to make a better assistant, they must be better than Apple on a number of vectors (too many to list, one big one that tech nerds often don't consider is Apple's retail footprint and the customer service advantage that provides).
Just curious, what are those things that Siri does and Google Now doesn't (except being more conversational and telling jokes, which 'Google Assistant' does too now with Pixel phones)?
I'm definitely bought over by the Google Assistant, it only improves upon the vastly useful and very smart Google Now already and is going to get a lot better with every release.
Hmm, I'll try that. Edit. Ahh, Siri seems not to know how to play albums, only songs. It also mis-parsed an album name of a song I have hearted in Apple Music. The mis-parse was for a more obscure title, but the incorrect behavior for playing an album was for a more mainstream one.
I've noticed that since my Apple Music trial expired I have to say "play MY song Self Control" or "play MY album Endless" for Siri to properly work. If I don't do that Siri just says it can't find the music.
Do you have an accent? Most complaints about poor recognition on google now comes from people with accents, it seems to struggle with what are typically considered heavy accents (some very localized accents or English as a second language speakers)
If not, try asking it any questions you can think of. You'll be surprised at how good it is, not counting "population of x" or "whats the weather?" type questions. Some examples: How long do you cook chicken for on the grill? When did the current pope become pope? Why do some people fill their tires with nitrogen? Do LED bulbs contain mercury? What's the loudest sound ever heard on Earth? Is it dangerous to drink distilled water? Why do canon printers have 2 black ink cartridges?
I use it a ton and it reads back an answer about 70% of time. Sometimes it misses easy ones and sometimes it nails really challenging ones. Sometimes it will work when your question is phrased one way but not another. Usually it prefers direct phrasing rather than vaguer phrasing (i.e. how long do you cook chicken on the grill vs how do you cook chicken on the grill). Overall though it is ridiculously powerful compared to other assistants.
I don't know which accent you're talking about, but I have a strong Indian accent, and one of the major selling points of Google Now is the amazing accuracy with which it works for that! Yes, you have to actually go and set the Voice language to English (India), because the default English (US) doesn't do that great for my accent - but as soon as I switch that, it works really awesome. When we friends sit together, we always compare our Indian English on Google Now v/s Siri. All I have to say is Siri is a joke in comparison. Not only does it listen to our accent correctly, it also speaks back all the regional words in the right way they're supposed to be spoken which is very comforting.
Lucky for you India is one of Googles favorite markets. They put a lot of effort into accommodating Indian users because of the massive untapped market. It's no surprise that it works well with an Indian accent.
The most common accent complaint I have heard is from people in the Southern US with very specific local accents. Look up Cajun and Appalachian English.
Google's voice recognition works extremely well, at least with certain accents. I'm from India, and it gets me right almost always. Not just that, it recognizes some ridiculously hard to pronounce names of places in India.
Fellow 5X user here. Seriously, I can't tell much of a difference between Now and Siri. They're both equally frustrating to use. Siri has the upperhand, imo, because it at least has a physical means of activating it (whoda thunk it...Apple with the physical button). Google Now I can long hold the "O" button, but then I still need to tap the microphone. There's always the hands-free "OK, -Assistant of Choice-", but that has always been incredibly hit and miss for me.
Siri is really bad. Another grievous example is once you choose a Bing result to explore from the Siri screen, there's no apparent way to get back to the results screen to try another one.
Stuff like this really confuses me. They went to all the trouble to have it say "sorry, I know what you want but I can't." Why not actually implement it instead?
Just mull the power and influence of deep learning: Yann LeCun is in FB, Geoffrey Hinton and part of his lab are affiliated with Google, Ruslan Salakhutdinov is now with Apple. In addition to this, there are other alumni of LeCun, Hinton, and Bengio dispersed across different notable companies like OpenAI, etc. Aside from the post-WW2 semiconductors-spurt, I can't think of a 'technology' that has become so suddenly so important as deep learning. You have a few scientists who spawned the fields of neural networks and deep learning in charge of what appear to be significant research efforts at the top tech companies (by market cap). Like silicon semiconductors and integrated circuits, deep learning approaches are likely to be the primary set of algorithms underlying many future 'intelligent' products and services. You will likely see a similar thing in biology/biotechnology with CRISPR in the near future.
I don't know who any of these people are. I am assuming they are all at the top of their field. Can you recommend some blogs or new sites geared specifically to AI news that you follow. Its a subject I am interested in. Thanks.
If you are really new to AI and want a broad understanding of the topic, I really recommend this book as a quick read: "Artificial Intelligence: What Everyone Needs to Know" by Jerry Kaplan
How about a book for a more technical, practical introduction?
I'm (hopefully) just starting a new position where some of the team is using machine learning techniques in an applied sense (i.e. it's not a CS position, they're not researching new methods, just applying existing ones). I'd like to get a good overview of common mistakes, pitfalls, etc. to look out for from a newcomer's perspective. I have used neural networks naively and briefly in a project about 6 years ago, but I know a lot(!) has changed since then.
A fair amount. I had a lot of math courses in university, quite a bit more than I needed for quantum mechanics and that sort of thing. But I'm also by no means a mathematician, and I'm definitely out of practice.
If your comfortable with calculus and linear algebra, then that is plenty to get started. Bonus points if you know probability.
Plenty of top universities put their ML material online, so I would pick your favorite and check it out. My school teaches roughly based on the Bishop book (I think he keeps a free pdf online). It's dense reading but has information on a huge array of topics. Someone else may be able to suggest other books that are a little more focused.
Really though, I would just pick a school, look at their course website, get their textbook, and work through the posted material at whatever pace you feel provides the value you are looking for.
Hinton, Bengio and LeCun are the most well known deep learning researchers. They wrote an introductory paper about deep learning for Nature and they also gave a talk at NIPS last year. That would be a good place to start I think.
There are not many blogs out there about the topic as far as I know, but I can recommend Chris Olah's blog.
What is really striking to me is that I encountered and read Hinton's earlier work in 1994 after coming across it in the stacks at the University. I was doing a lot of quantitative social science research in my grad program.
Once I left the university and started doing market and customer research in business that kind of raw prediction/black box modeling didn't fly very well. Companies preferred simpler models with explainable connections between independent variables that they could leverage and dependent variables (business outcomes).
Smart move by Apple. I'll be following their moves to improve Siri closely. The fact that they're hiring researchers as strong as Ruslan makes me feel that they've got real potential to drastically improve Siri going forward.
I think that Siri has the potential to improve non-linearly. If they can get Siri to be slightly more useful, the millions of iPhone owners will instantly use it more, creating a virtuous feedback loop.
It's not clear if the acquisitions they've made in the past have been integrated into any of their products yet. It's certainly within the realm of possibility to wake up one day and use a new and better Siri.
I would say almost all of Apple's acquisitions are acquihires and for intellectual property but I don't see how that has any relation to how long it takes to integrate the people/tech into existing products. My guess is as good as the quotes in the article.
It sounds like what they're saying is that DNNs are used for their speech recognition (speech is split into acoustic and language modeling, and DNN speech models have been state-of-the-art for about 4 years). It's kinda a "so what?" that they were using DNNs for their acoustic models, and nothing to do with the interesting parts of the system.
If they were using DNNs for their NLU system back then I'd be kinda surprised.
Can you point at a good survey article, blog post, course or book that goes into the NLU aspect? I've watched a bunch of courses on DNN but feel that stuff doesn't apply well to NLU aspects. The speech rec stuff used to be done with HMMs until recently but as you said, DNN is now the way to go for it.
(also, sorry, the speech stuff was using HMM+GMM models, and now mostly use HMM+DNN models [so, we replaced the gaussian mixture model with a DNN, but kept the HMM], although some people are moving over to RNN models which doesn't use HMMs (connectionist temporal classification))
an intent classifier to predict intent (e.g., "play music"). This can be pretty much anything (SVM, logistic regression, RNN, I don't care, try stuff and see what works for you)
a segmentation model or slot-filling model (so label every word as either junk or "album name" or "song name" or whatever). This requires a sequence model, e.g., HMMs, CRFs, LSTM-RNNs.
https://www.cs.toronto.edu/~hinton/science.pdf