I would argue that we might be in a Research bubble(who knows.. I don't have a clue) but we are definitely not in "using the ML for practical problems bubble".
Amount of industries and businesses that can benefit not from the state of the art but from stuff that been known for years now and just works because it was developed for harder problems is huge. The big return/promise companies are vacuuming up all the talent while niches all over the place can benefit from someone spending few month cleaning up data and using some transfer learning to save them a ton of $$.
For these application you don't need PhD you need an engineer who knows how to ship stuff but on other hand also knows how to work with current DL frameworks.
I’m less skeptical about this than I was about the blockchain bubble, but I’ve yet to see AI or ML actually work on a smaller scale.
It’s a big hype in the public sector in these years, but it’s really all talk. We did a project where we used ML and 1000 server instances in Azure to go through millions of employee cases, to flag cases that didn’t have a certain document. Because the the is new, and this wasn’t something that could be delayed, we also had 10 employs do the same task to make sure it got completed.
The ML project took 2 really expensive employs and 3 months to train the algorithm, then 5 hours to go through all the documents + we had to spend 1 week of 5 employs to clean up the stuff our algorithm had marked as “unreadable” due to terrible scans. The human employs did it in the same amount of time, but made a few more errors.
Over all the ML was more expensive and our politicians won’t favor it again.
On other things where ML might work, our datasets are turning out to be too small, unless we work together with other municipalities and then we’re facing GDPR violations that may be hard to pass through our legal team. Legal is a big issue on a lot of things, it’s not currently legal to automate any sort of process that require any form of validation (as long as it has to do with case working).
That being said, I’m sure stuff like facial recognition will change the world.
As a single batch job that won't be repeated this doesn't sound like a good candidate for ML. ML is more suited to on-going processes.
Why would you use server instances in Azure to do ML? Something like Google CloudML (I'm sure that the other major cloud providers do managed Tensorflow as well I've just never tried it on their platforms) would be a better fit to a project with only two technical staff. Your two staff probably spent a combined total of one-person-month working on infrastructure.
Your issue with small data is very real. People need to stop trying to do ML on small datasets. The results will be sub-optimal.
When you say ML, you must mean deep learning applied to unstructured data (vision, audio).
ML in general can absolutely be used with small datasets. ML is all about finding the right model complexity to fit to the data to maximize out-of-sample performance. If your dataset is small, all that means is that your model will have to be more crude. A simple cross-validated regularized linear regression or a shallow decision tree are ML models too, and you can usefully apply them to a dataset of just 100 samples.
It depends on the application. You might want to use the data to verify your intuition, which may not be consistent with the data. Or your intuition may have been based on the very same dataset but overfit to it.
I do think that people that think that ML = big data are mistaken. ML is about making the most of your data, however much of it you have.
ML is ill-defined. But take most textbooks on ML, and you will find regularized linear regression, decision trees, cross-validation and bootstrapping all in there.
In my view, the main difference between ML and plain statistics is that with the latter, you come up with the appropriate model apriori, and then make sure the data satisfies the assumptions so that you can draw conclusions from the in-sample fit of the model. You control for overfitting by choosing the simplest model that is reasonable - often univariate linear regression.
Whereas with ML, you let the data dictate how complex a model you should use. You choose the appropriate model complexity using techniques like cross-validation, and verify the effectiveness of your model empirically.
ML is often used interchangeably with ANNs which I think is a mistake. Take structured data problems on Kaggle and you would very rarely see ANNs as a major predictor in the winning models.
Not necessarily. Cross-validation can give you a valid estimate of out-of-sample performance even if you have more dimensions than samples, and even if some of the features you have are (sporadically) perfectly correlated with the target variable. See https://stats.stackexchange.com/questions/295626/does-cross-...
The terms you're using - they come from Machine Learning, which comes from the same division of mathematics as ordinary statistics. And that's how ML is educated at the universities anyway.
So if you've got an analyst sitting on that issue - then your problem is solved anyway.
So again, why hire somebody with a trendy specialty who is <probably> full of BS, when you can hire someone with some respect for classics?
Every year someone releases a new altcoin with a youtube video that promises the end of war, starvation, etc... Then suckers buy it. 3 years ago, it was developing 'the ultimate privacy coin', because the libertarians ate that shit up.
Using blockchain outside of currency seems expensive and unnecessary.
I dont see bitcoin bubble popping unless they have a security issue. The rest of the coins might be as good as useless when people realize how expensive it is to run 7 computers to do the job of 1.
There is a lot of old tech out there that could be implemented and would probably provide ROI, but isnt. So many businesses are still running on paper! My point is that industry isnt just sitting around waiting for new tech so they can get higher efficiency, there are reasons why not everyone is doing the latest and greatest (including ML) and they're usually valid.
This sounds good on paper but I don't feel we see all that many real examples of this sort of model (engineer + niche problem) or at least haven't heard of them. But then again if we did I guess that opportunities already gone or saturated.
Different areas of manufacturing for example. All that computer vision, now apply it to X-rays of metal parts used to try and detect failures(not it is a guy looking at them all day using a dozen primitive filters, 8 hours a day every day). These are very domain specific, like liners of internal combustion engines, turbine blades etc
On app side.. Cookpad added this neat feature last year, where it scans your pictures and adds to you cfood/cook log if it is food. The had a PhD do it, but this is because they want to do a lot more with it later I guess so building expertise and team. Food in general I think has a lot of neat computer vision apps that will happen eventually.
A few things, or rather ripples, survived the bubbles. AI crashed in the 1980s - but we eventually created Dragon Speech, handwriting recognition and a few other useful solutions. The DotCom bubble burst, but we wound up with [arguably] useful results in the long run (agile, or whatever comes next, might be the biggest ripple).
Even if this bubble bursts, bigger and greater things will arise because of it. With only 2 samples of these bubbles the p-value is miserable, but history has taught us to see these bubbles through.
Wisdom is not the sum of success, it is the sum of failure. We need to learn just how far we can push deep learning - within reasonable moral degrees of freedom.
Bubbles (last I checked) don’t have a firm empirical definition. But the ebb and flow that you’re arguing for is very dramatic. Society could have made those discoveries without leaving large swaths of people and places unemployed. Further, anything that impacts the populous negatively makes a bigger, deeper, and longer lasting, negative impact on national wealth than anything a few companies could ever achieve in similar constraints.
> Society could have made those discoveries without leaving large swaths of people and places unemployed
I am currently a proponent of self-driving cars, as I believe that human reflexes are designed to operate at, on average, 6mph. Almost anything could be better than us, at least in theory. Testing the limits of self-driving cars is how we determine what we can't do with them (and asking too much is how we zero-in on the failure point). This has complex moral consequences. Will more people be killed due to failure, or fewer? Thats great if it works out, but horrific if it doesn't. It's incredibly difficult to draw a line given that we don't know what the whole landscape is.
This bubble is worsening unemployment. One ripple arising from that is UBI. Another is a positive-sum world. ML is currently successful in medicine. It's not all doom and gloom - we are already reaping long-term benefits.
Our reflexes are fine at about 100 Hz. That gives up to 2m distance at rival urban speeds before action and drivers shouldn't rely just on reflexes anyway.
The problem is not reflexes, the problem is some drivers being dumb and a bunch of others distracted. You could mandate a very limited systento prevent both without solving the self driving car problem at all. (I recommend looking at trains for this solution.)
It is being done because it is attractive, not because it is necessary.
By the measure of the parent comment, and, I think globally. It's often best to accommodate an argument when you have more than one way to make your case. If someone believes that unemployment is on the rise, despite the facts that you (specifically) may have, let them win it - because they have seen the same facts and don't care.
The DotCom bubble burst, but we wound up with [arguably] useful results in the long run (agile, or whatever comes next, might be the biggest ripple).
The outcome of the dotcom bubble was the build-out of the infrastructure (dark fibre). A similar bubble built out the railway infrastructure the century before that.
I would say failure by definition cannot be held up as a goal.
The upside of failure is that it teaches more reliable lessons than success, there's no survivorship bias at play when learning from failures like there is with success. Jack Ma's lengthy interview at Davos this year touches on the subject if you'd like to see that perspective defended by someone with authority.
> I would say failure by definition cannot be held up as a goal.
It depends on who is doing it. Maybe you understand what you said, but others have an interest in not understanding it because they have an obstacle in themselves.
> The upside of failure is that it teaches more reliable lessons than success, there's no survivorship bias at play when learning from failures like there is with success.
I would like to improve your abilities or otherwise open up your wisdom. Any "upside of failure" is by definition a success. Therefore people's statements about success and failure, as if the nature of the result can be captured by a singular datapoint, are self-contradictory, and meaningful only as a reflection of the individual's limitation of understanding.
> if you'd like to see that perspective defended by someone with authority.
We're talking about wisdom here, so the only authority is a perfectly enlightened being, not an ordinary human being who has knowledge but no way to confirm the degree of exactness of their own ability to make confirmations. No one without enlightenment perceives the causes of why things happen to them. If you want to convince me someone is an authority on how a good and a bad thing happens then I will need to see any proof at all they realize the principle by which the world operates. Otherwise you should realize that's merely their own ideas, of which some percentage may be correct, despite their inability to confirm it.
They were simultaneously different points and part of the same point in my argument. Separately, failure is a good thing - as any CEO would tell you. Together, they are horrific.
It's very complicated, which is a far cry from "idiotic" as the article portrays.
Go ahead and say what they are if you know them. I doubt you will be able to say anything. Instead of being silent, you should have the courage to admit you don't actually know what you're talking about and don't have much if anything to tell me.
Except when you are being taught by another person, all new knowledge comes from your own observation, trial and error. When you are learning from another person you are learning about the aggregate of incremental failures that they made, finally leading up to a success. Very rarely do things totally succeed on a first attempt.
How can you possibly say failure can't be useful? How do you iterate your plans and designs? I don't think you appreciate how knowledge works. Your weird demonization of failure (including it somehow being a moral issue?) won't force success, it's just going to stop you from realistically evaluating your goals.
Aside from that explanation, I'm not dying to debate you or list exhaustive examples or whatever because you're belligerent and arrogant. People like you can't be convinced, you have to go through it yourselves several times to figure it out, if ever.
You need to double check your interpretations of what I said. If you fail, you fail. If you treat it as a good lesson and learn something from the result, is that the failure making you learn or is it you?
Aside from that, you have described yourself in the remainder of your comment's criticism of me. Go ahead and check out of a real conversation with me even though you don't know a single thing about me. You've already done all the damage. People in the present might be fooled by your pretense to humility and ability to diagnose arrogance, but in the future people will look back on our comments and find it very easy to discern that you don't even know yourself.
Right now most AI experts in the industry are financially incentivized. They are more passionate about the money that they can make than about the research. It attracts the wrong kind of experts.
The real experts are those who are working for universities and getting paid peanuts. Those who value knowledge above money.
There are "real experts" in industry too. Industry doesn't need only good algorithms with solid foundations, which are invented in universities (say, SVM), but also people who understand the domain and can code that knowledge into algorithms and data (e.g. feature engineering).
I get the sense reading this article that the main concern raised is whether or not enough value is being generated by companies and researchers that "use machine learning to solve problems related to X" to justify the investment. The author doesn't touch upon whether using ML could be exacerbating the very problems these implementations are trying to solve. Inappropriate implementations of classification algorithms and predictive analysis that create false positives has the potential to inflate this sort of a bubble even further.
If this bubble is to burst it will be followed by a "trough of disillusionment" (see the gartner hype cyle). But then hopefully the "slope of enlightenment" - unfortunately many of the early companies won;t survive the disillusionment phase.
:-) Point taken. It was first invented after the 2000-2001 dot com bubble I think. However i think it has some merit. Some of the things that were part of that period's hype (eg migration of retail to ecommerce) did . Eventually (nearly 20 years so far)....
nah, there is no bubble. We just need to combine it with blockchain technology and then we have something shiny, rock stable which will change the industry ;)
Amount of industries and businesses that can benefit not from the state of the art but from stuff that been known for years now and just works because it was developed for harder problems is huge. The big return/promise companies are vacuuming up all the talent while niches all over the place can benefit from someone spending few month cleaning up data and using some transfer learning to save them a ton of $$.
For these application you don't need PhD you need an engineer who knows how to ship stuff but on other hand also knows how to work with current DL frameworks.