In the next decade it will start taking over computer graphics, medicine, manufacturing, surveillance, hardware / software and every other aspect of our lives.
EDIT: for people who are downvoting me, here are some examples of how machine learning is and will transform our lives:
1. 3D graphics: https://www.youtube.com/watch?v=FlgLxSLsYWQ Andrew Price doesn't have a technical background but does a decent job summarizing some of the ways deep learning will transform computer graphics. Soon all of the mocap, character design and animation will be driven by deep learning systems.
2. Computer Architecture and Traditional Software: https://www.youtube.com/watch?v=TTpKWOuzOxc There's a ton of recent research showing that you can use machine learning to beat human crafted heuristics in hardware, scheduling, compiler design and query planning.
I would say it's table stakes for large companies in certain areas, but not small (see my other reply).
Machine learning seems to be good at squeezing percentage points out of monopoly with a business that already works. That's very different than "table stakes for small companies". Small companies need optimize their product with fundamental changes.
For example, in the early days of AirBNB there was a more inconsistent user experience, which you heard about in the media. I think they addressed those problems with policies, incentives, and some elbow grease (kicking bad users off the platform, etc.). Not machine learning. Machine learning doesn't help you expand into Europe or Asia, etc.
Yep, and the results are hilarious or tragic depending on how you look at it:
We keep getting ads for the thing we bought yesterday, ads for dating sites after we got married and had kids (continously, for ten years despite my utter lack of interest), search results keep getting worse[0], obvious spammers keep on spamming in social media (seriously, it seems a simple regex filter could have done a better job to reduce crypto scamming in replies on Twitter than whatever was there last time I checked.)
[0]: some people will always claim it is because black hat SEO is so much worse, but that doesn't explain why Google sometimes can neither understand doublequotes nor the verbatim option anymore. That is not the result of black hat SEO but of sloppy maintenance, and I guess so is a number of other problems.
That's because in a lot of cases they're optimizing for the wrong metrics, as in maximizing their revenue instead of your utility.
There's way more content on the web than there was in the early 2000s, most of it in form of "content marketing" and explicitly attempting to game the system.
If you look at recent results from TREC, it's pretty clear that machine learning provides a large boost over the traditional retrieval systems, on any metric that you want to optimize.
What? How do they maximize their revenue by spending money on ads the users actually laugh about for how bad their targeting is?
Based on what I know about machine learning - it almost always gives great short-term results, and it almost always fails to deliver the expected long-term results. What's worse is this comes with the weirdest most indecipherable bugs that pop up more and more over time. Unless you have a large enough database to show statistical errors that are negligible (something like 99.9999% precision or recall, depending on your metrics), you should assume it will break in ways you cannot possibly predict. And even then, you might be using the wrong training data without even realizing it.
I'm not saying ML is bad, although I am saying it is ridiculously overhyped. I'm saying ML is still nascent enough nobody really knows how a lot of edge cases will shake out, simply because there are too many edge cases to test before putting it into production.
It's not hard to find examples of ML algorithms gone wrong even for sites like Amazon.
https://gizmodo.com/amazon-prime-day-glitch-let-people-buy-1...
You can have a system that generates them a ton of cash while making mistakes in some cases, outliers are inevitable and feedback loops in recommender systems can lead to such issues. Amazon wouldn't deploy these systems if they didn't move the needle.
> Amazon wouldn't deploy these systems if they didn't move the needle.
This is an appeal to authority that Amazon executives are immune to making mistakes. They could have bad metrics. They could have bad incentives encouraging managers to make poor decisions (basically this describes everything wrong with Google today). They could be incompetent. They could be focused on short-term gains at the expense of long-term gains because it maximizes their personal net wealth and they can just jump ship in a few years.
Again, nobody is saying ML algorithms are worthless. I just don't believe (and have lots of reasons based on personal experience I won't go into) that they are 10% as useful as the industry wants you to believe.
Anyway, agree to disagree.
There's a story, and I think it was (re)posted here recently about a series of MBAs all optimizing the cost of the burger bums by removing seeds until there are three seeds neatly laid out at the top.
It might be measurable all the way but at some point it becomes ridiculous. For me that time was some months ago. For the rest of Internet they might manage to reduce the quality once or twice more before it becomes obvious.
This is my way of fighting back. By posting here and on my blog and getting upvoted a lot for pointing out what many can already feel. By letting people who read HN know that yes, that feeling they have that a lot of their ads are wasted because of bad targeting might very well be true.
I'm against most forms of recommendations as well, that doesn't mean that they're not valuable to the businesses deploying them. In most cases the end users of these systems are not the real customers of these platforms, and they're there to serve the advertisers who keep these businesses running.
Not if it's something I only buy once a year, for example. That's where "learning" part should come in. You don't need any learning to just parrot me back my inputs.
> Amazon wouldn't deploy these systems if they didn't move the needle.
I don't know about Amazon, but I've recently read on HN some articles strongly suggesting almost nobody is properly measuring the impact of ads, let alone the impact of "targeting". In most cases, people more or less just stuff money into ads budgets, because that's what you do, and they get customers - because people still need to buy things, regardless of any targeting - but the casual link between the former and the latter is not really very well established.
I am not sure how these are contradictory. My interest is buying things I want. Their interest is selling me things I'd buy. How showing dating site ads to a married person promotes any of those? Where the revenue would come from? Or do you mean it's ad agency revenue, not advertiser revenue? In that case we clearly have a case of agent problem.
This does not demonstrate that the machine learning algorithms are doing a bad job. Quite the opposite really. See:
As for ad placement, social media, and search ranking: is that actually effective? The one constant with ads and social media and web search, IME, is that it just never gets any better. From the Lycos/AltaVista days to Google there was a huge jump in search result quality, and then it plateaued.
The ads I see online are still laughably bad. They're more professionally produced than 10 or 20 years ago, but they're still never for any category of product that I'd ever buy. Targeted and tracked advertising seems like a big scam.
Those were just some of the most visible examples. OCR, image recognition, NLP and speech recognition alone are enough to revolutionize most data entry jobs and we'll see a ton of startups using them to do just that for every industry.
There are countless other examples of machine learning applications, including drug discovery, radiology, production line quality assurance, etc.
> The ads I see online are still laughably bad.
Have ads ever been good? You see the ads that you're seeing because someone is paying a lot of money to put them there.
Bloomberg spent 120 million on ads last month, ad platforms had to find eyeballs for them.
If I'm looking at, for example, "how to programmatically control my model railway with a Raspberry Pi", there's a large number of easy to target, relevant products you can advertise against that. You don't need a gigabyte of creepy backstory on me to figure out "hey, show me ads for the products mentioned in the article, and there's a good chance I'll buy them because it's convenient."
Yeah, Mike Bloomberg might be willing to pay $1 to show his ugly mug next to that page, but the ad from a modeler's supply shop, who would only have paid 95 cents, would have felt much more appropriate to the consumer. I know some ad networks try to measure quality via clickthrough rate, but I'm not sure it was weighted in a way to nerf this issue.
Another problem is that a pure-content model doesn't extend to all types of site. There's no suitable product to advertise against "six die in New Year's party gone horribly wrong." and nobody wanted to run low-yield CPM branding ads, so they had to backfill wth retargeting and profile-based ads instead.
There's still a lot of anti-data sentiment in the paid ad industry, where media buyers will guide themselves by what they think their customer looks like, and not what data tells them it is.
And until this is a fixed behavior, you'll keep on seeing untargeted ads.
I have bought many things from ads related to the media I am reading. For example, I read muscle car magazines, and buy parts/tools from the ads in those magazines. Back in the era of print mags, my company would do well placing ads for compilers next to relevant programming articles.
The counter-example to your point is that deep learning maniacs were predicting self driving to be a done deal by 2020 and we now realize we are so far away from that goal. Machine learning will only work well for one-pony tricks problems that can be clearly isolated. Nothing like "every other aspect of our lives".
And the results are, as far as I can see, horrible. I mean yes, if I make a mistake to search for shoes on Google, half of the internet will be showing me shoe ads for the next 6 months (how many feet do you think I have? Do you think I buy shoes in dozens?) - but presenting it as the triumph of ML is IMHO an overreach.
> personalization on all top social/media apps
From what I see, the same apps struggle to not get sued because of the said personalization regularly pushes sex content on kids, triggers on snowflakes and election interference on potential voters. Or at least that's what I hear every time next round of censorship is introduced into the social media. Somehow I am still not seeing a cause for celebration here.
> search ranking for google
Same as above, plus one shouldn't use Google search anyway. Use DuckDuckGo.
> translation, speech recognition
OK, here it got pretty good results, though sometimes it is as good as a very drunk chimp who got into a dictionary store, but in many other times it's decent. You get this one.
> finance, fraud detection
As a consumer, haven't noticed it. 100% of fraud on my cards have been detected by me reviewing my credit card statements. 100% of fraud alerts by banks have been false positives. I do not begrudge that specifically, I'd better have false positives than more fraud, but not seeing much ML-driven progress there tbh.
> In the next decade it will start taking over computer graphics, medicine, manufacturing, surveillance, hardware / software and every other aspect of our lives.
Not sure what "computer graphics" means, medicine probably not, surveillance maybe, but that's exactly the opposite of what I'd want, software definitely not even close, for the rest I'm not even sure what you're talking about.
> here's a ton of recent research showing that you can use machine learning to beat human crafted heuristics in hardware, scheduling, compiler design and query planning.
I'll believe it when I see ML-driven code generator doing something useful without tight supervision by humans. I can believe ML doing specific heuristics better (heck, doing exactly that has been part of my job recently) but there's a huge difference between "figure out exactly how much sugar makes the specific cookie recipe taste the best" and "invent whole cookie recipe from scratch and bake the cookie". I am sure ML would be useful for the former, for the latter... I'll believe it when I see it.
Having worked in this industry for a while I just want to point out that the vast majority of fraud detection is done for the merchant, not the customer.
Pretty sure ML have shown promising results in cancer classification in images (such as xray, etc). Not sure how this will pan out but the limited scope seems ideal .
I do think in general Machine learning is only just started with respect to All industry, not just tech. There is so much shown what could be done, the next 10 years will be very interesting.
Machine Learning has made amazing strides in the past decade and will change all aspects of our lives. There will be a ton of startups in the next decade using speech recognition, image recognition, natural language processing and etc to automate all of our repetitive tasks. If hardware keeps up we'll probably see a boom in robotics in the next 5-15 years that will allow us to automate most manufacturing and food production.
AI is a branch of computer science, but outside of A* search it's practically all machine learning.
I believe the demand for companies to actually use the fancier parts of machine learning are going to be much fewer than people think. Number 2, on the other hand, is much more generalizable, and is having a much larger effect on the 'average' business than NLP and image recognition.
I used to work at an image recognition API startup and went through an AI incubator with my own company so I got to see the huge range of applications of this technology. At the incubator we had a company automating customer support (acquired by google), another one automating appointment booking and scheduling, sales call analytics using speech recognition, automatic reading comprehension quizzes for K-12, deep learning for radiology, patient tracking/monitoring at retirement homes and hospitals, imitation learning based robotics, deep learning based hedge fund, amazon go like vending machines, fashion analytics, social media analytics/engagement for brands.
I use my Google Home Mini every day and it's just like the computer in Star Trek. If you could show it to someone from the early 90s they would certainly say it resembles an AI system.
The computer in Star Trek is just a smart speaker. Lt. Cmdr. Data is AI.
I think the usual terms we use nowadays is narrow AI and general AI to differentiate both types.
What are some examples of companies that hinged on machine learning?
DATA is crucial, but that's very different than "sophistiated machine learning models". For example, do Uber and Lyft need machine learning?
As far as I can tell, they need data about drivers, users, ride completion, maps, and a lot of people doing simple statistics on that data and improving the product. Machine learning might give them a few percentage points in some specific subsystem, but it isn't going to make or break the product as a whole. It's not "table stakes".
Pick another industry and do the same thought experiment. What are some other breakout companies? Did snapchat succeed (in its day) because of machine learning? What about AirBNB? I think they are optimizing the product with "data science", not machine learning. Certainly not deep learning.
Machine learning finally came of age in the 2010s and is now table stakes for every tech company, large and small. Accumulating a data asset around your product and service and using sophisticated machine learning models to personalize and improve your product is not a nice to have.
My favorite is that they do pothole detection, prediction and routing around based on what they classified based on driver accelerometer data. They've also reduced traffic incidents with insights from the same dataset that drove product changes.
To me, the accelerometer DATA is the new and important thing -- it's the thing that taxi companies don't have.
Whether the machine learning model is "sophisticated" doesn't seem critical on the face of it.
I guess it depends whether there are a lot of non-pothole events that look like pothole events to "naive" signal processing. Do you need deep learning for this problem? I don't know but I would be skeptical without a citation.
But I am interested in more info and examples. If someone can give a bunch of examples I would adjust my view.
Data is obviously crucial, but it's useless if it's just sitting on hard drive. Machine learning allows them to squeeze out as much juice from it as possible, and a few % point lift makes a huge difference at scale, especially in high volume businesses with low margins per transaction.
More importantly it's worth understanding that ML allows us to do things we couldn't do before, because it's cheaper and scales better than previous optimization methods, when you have the data.
So yes, the data infrastructure is the critical part, however the ML pipelines and Data Science work is what gives the massive multiplier at a cost/scale you can't get with traditional optimization methods.
Alexa / Siri / Google Assistant would not be possible without machine learning. Same goes for Snapchat's filters and any other form of AR. Google's Pixel phones wouldn't have such a great camera without computational photography / machine learning.
That said, uber and lyft certainly benefit from much more interesting ML and stats than just "simple statistics." For example, setting prices and incentives to properly balance the driver side against the rider side of the marketplace.
And yes I probably should have said "data science" or "statistics" rather than "simple statistics", definitely for Uber and Lyft (less so for smaller companies).
I worked on a data science team 10 years ago when it was called "statistics". I even remember a 2009 NYTimes article that talked about data science without using the word "data science", because it hadn't come into use yet.
If people want to redefine "machine learning" to mean "data science", OK fine. I guess that's what's happening now. The term "data science" already annoyed some people, but I guess now it's "machine learning" and "deep learning".
Anyone doing large scale ETA prediction and routing these days is using ML models at various phases in the pipeline, from GPS noise filtering to road network data generation. You might get past the early stages of a rideshare or delivery service with more traditional naive models, but there are still significant efficiency gains to be made in even the most cutting edge systems today. For maps and navigation, this is especially true when you go global and need models that adapt to subtle cultural and behavioral changes across markets.
Real-world win conditions tend to be more complex. For non-general AI tasks, this is much less of a problem. "Identify hot dog vs. not-hot-dog accurately 99% of the time" is measurable and verifiable. But a simple goal for a complex general-AI task is likely to build a Paperclip Optimizer. I suspect we don't even have the right language to build a good win condition for a lot of real world problems-- we may end up ignoring "insignificant" factors only to discover they're huge over tine or at scale.
Games also tend to provide near-perfect or perfect information. At a minimum, there's constrained state transitions. I can't pull out my Oyster card in the middle of a poker game, add it to my hand, and declare it's actually a four of clubs. Real problems can have surprise off-the-wall, or outright externally triggered transitions, that completely throw the model off.
"AI, build me a market-beating stock portfolio to cash out 2040-01-01" is a lot harder than "AI, checkmate this king."