Well, I can understand why the author would come to this conclusion. A 3rd-year undergrad being called by companies and VCs for insight? I'm impressed by the author but disdainful of these solicitors. It reflects poorly on them. To me they seem like people who don't understand what's happening but are anxious not to miss out on a new wave of get-rich-quick schemes.
Like other commenters have mentioned, largely due to misbranding and sensational media hype. Fearmongering from people like Elon Musk hasn't helped. But the key impact of machine learning for me is to make better and more efficient decisions that are informed by data - and that is not going to go away.
>but disdainful of these solicitors. It reflects poorly on them. To me they seem like people who don't understand what's happening but are anxious not to miss out on a new wave of get-rich-quick schemes.
This is an odd point of view to have. While op may see this as pressure because of imposter syndrome, I would have DREAMED of an opportunity like this in college.
These entrepreneur may not quite understand what they're getting into, but make no mistake, they are very good at recognizing experts, both established and rising [although they do cast a wide net] and this kind of behavior drives life changing opportunities and make it substantially easier, in a difficult world, for the cream of the crop to be unlocked to exceptional success at an early age. And the rising tide helps all of us.
The smartest people in college are usually the ones who ate able to learn what they need to know outside of college on their own, once they have an income, and entrepreneurs unlock that potential.
I can see your point of view as well... I would've been pretty excited to have this opportunity also.
However call me idealistic, naive, whatever - I don't like the motivations of these solicitors, even though in a pragmatic sense their actions will have positive side-effects for us (the "rising tide" you mention).
No offense to the author. But serious question: why whenever articles like this come does everyone say "oh you're not a fraud at all!!". You know what, maybe she is a fraud. If she's being honest in this post she sounds a bit like one. She's crafted the perfect resume to attract attention about AI and it works. But she doesn't really know all that much.
Possible that she has severe imposter syndrome and she sounds above average. But maybe she really isn't as great as everyone thinks she is and she wants people to know that. And maybe people shouldn't pat her on the back and say "no dude you're great don't say that"
*disclaimer: basing this on blog comments and some comments here
Also, there's something I don't get: how does an undergrad ends up teaching a full blown course at Stanford? Especially considering she doesn't seem to be some kind of super genius, only a probably excellent student. I'm not trying to demotivate her, I actually think it's amazing that he/she is doing this, but how did this happen?
I have the same background as the author, but a few years older.
Stanford undergrads and grads alike can teach a course as long as they have (1) the necessary background, (2) passion and proficiency in the tech, and (3) motivation to teach and manage course overhead.
Being a "super genius" doesn't correlate with being informative and having intrinsic instructional value.
I totally agree with your last point, and I wish more university acknowledged this when deciding who teaches classes. My favourite CS course was taught by a first year MS student. For most universities though, letting an undergraduate teach a class is a no-no and it surprises me that Stanford, one of the most prestigious university in the world, would make an exception to that.
Edit: Commenter above clarified to me that the student is not teaching a full blown 3 unit course.
The author seems to be teaching a "student-initiated course." Many universities give students to teach a 1-unit course that is not really part of the mainstream curriculum and does not really count toward the degree.
Same happened to me on a big European technical University. I pushed for adopting a new course, I was the world expert in it, I thought the class for 10 years until someone in the government pushed for a real professor. Now they have another guy who is much less qualified than me, but I was very happy and successful in the private industry instead.
And in the end I thought in 3 different universities.
I don't know about Stanford in particular, but my experience so far has been that if you just volunteer to do stuff (and maybe have rear cover from one higher-up, but that's often not really required), and don't need any budget, people just let you do stuff.
At work, I started a study club, and started teaching a programming course, all on company time. No pre-approval. Nobody objected.
And the universities I've been to tend to be even more liberal than companies in such matters.
So an undergrad teaching a course doesn't seem questionable to me.
It doesn't really matter if the OP (or anybody else really) is a super genius or not. People from outside the field can't tell, and so they approach those with a high media profile, and who match some of the preconceived ideas, like the buzzword bingo mentioned in the blog post.
The same is true even inside our field. When I'm interested in some architectural design pattern, I often read an article by Martin Fowler. Why? Because what he writes sounds plausible, and because I've heard his name a hundred times before. I've never seen production code he wrote, or been on a project he worked on. Maybe he feels like an impostor too, sometimes?
I'm not calling Fowler an impostor; I just want to draw the parallels how second-hand knowledge influences our perception of expertise.
The simple answer is that this blog post tells us nothing about her skill level aside from the clear resume brags. Google I/O presenter on Tensorflow said something along the lines of "why do we use RELU? Because it works better.. Why does it work better? We don't know for sure." Any field where the basic questions are still unanswered is going to make talented people think they are posers. In fact it's the only rational stance..
Agreed, but she exploited their algorithms which is definitely something 'frauds' do. I'm not even saying it's a bad thing, it's working for her. She shouldn't feel bad but she shouldn't feel like a genius either.
> Even though I’m one of the beneficiary of this AI craze, I can’t help but thinking this will burst.
I don't think it will. Level off - maybe.
I've started my work in Computer Vision with classical algorithms (SIFT features, geometry, correlation filters and things alike people were researching for decades). These really worked like garbage, it was a nightmare.
Then we jumped on DL bandwagon - and CV just clicked for me. Now I see it working, not perfectly, not at human level yet, but it works, it's better than everything else and it certainly brings value - not just in CV! Maybe there will be some expectations delayed or even ruined (AGI, fully self-driving cars, dunno), but the tech isn't going anywhere.
At it requires at least some experience and a specific mindset, slightly unusual for a generic programmer. So I don't see a problem with experts, courses, degrees and the like.
Pattern matching is the one thing DL is good for. Which is why it's a good match for CV. Calling DL AI in the first place was a mistake or at least over zealous marketing.
Playing go or chess or matching patterns are all things intelligent begins can do but that does not imply that doing those thing means you are intelligent.
One can argue that the parts of our brain that make us intelligent, the prefrontal cortex that is so much bigger than in "lesser" animals, is essentially an overgrown, glorified pattern matching engine. Pattern matching is the one thing our brains are good for - there's good reason to suppose that quite many intelligence-related tasks can be reduced to a form of pattern matching.
One could argue a lot of things. Humans once argued quite seriously that the human brain was composed of microscopic gears because that was the technology of the time.
"Fairly" according to whom? One of the nations? Some individual? Or do you want the AI to come up with the definition of "fair"? What if some people disagree? What if the majority of people disagree? In any case, this could be achieved with extreme violence. Do you really want "AI" imposing its (or someone's) will on the world?
Not even at insect level yet. There's no doubt things will improve, and there's already great value, but I hate calling ML "AI". It's been over 70 years of ML research (specifically neural networks) and I don't know how long it's going to take to reach insect-level behavior (which is still far from basic intelligence) let alone so-called AGI (which, BTW, people in the '50s were certain is just around the corner), even though I think we'll get there eventually. We'd better stop using the term "AI" to mean anything other than a field of research or an aspiration, and definitely stop using it to describe existing software.
Computers were able to perform some tasks better than humans since they were first built in the 40s. Computers are only used for things they are better at than humans. There is nothing to suggest, however, that there is anything closer to "intelligence" in recognizing images as in compiling census statistics, and no software is as adept in "general problem solving" as insects.
I'm sure you're familiar with the line of reasoning that if you asked someone 50 years ago to describe tasks that require intelligence, they'd for sure say recognizing objects in images is one of them. Now that computers can do that, it's no longer 'intelligence' and the goalposts get moved.
In what sense is no software 'as adept in "general problem solving" as insects'?
People also characterized doing arithmetic as intelligent, and there's little doubt that the entire idea of calulating machines -- from the time of Leibniz -- was motivated by the desire to emulate the human mind, so there's no point in describing intelligence as a binary quality. Therefore, I don't need to define what intelligent means in order to demonstrate my point. There are cognitive tasks that insects do better than computers, ergo, we are not yet at insect-level cognition. Only once computers can do everything better than an insect can we claim that they are more intelligent than insects.
Humans can't do _everything_ better than insects, does that mean insects are more intelligent than humans?
Chimpanzees have better short term memory on certain tasks than humans [1] - humans not being better at _everything_ than chimpanzees doesn't make chimpanzees more intelligent.
You are being very literal. That people are amazed that chimps are better at some particular task is exactly the point. It's the exception that proves the rule. When we're amazed that insects are better than computers at something we can start arguing over which is smarter.
This. I don't think people even get a hint of what is possible nowadays with DL in CV, NLP etc. I am actually depressed when I talk to some friends and they are so pitifully outdated, and then even after showing them how to do some magic in 100 lines, observing they are still not getting it and continuing in their old ways :(
Just out of curiosity, where would you recommend starting for those outdated people? Machine Learning, Deep Learning, AI (for lack of a more specific acronym), NLP - these things are kind of daunting for newcomers, if only due to the acronym du jour changing constantly.
It's difficult to say to be honest; for me the "enthusiasm" works best, I simply picked an area I wanted to know (e.g. self-driving cars using DL) and then learned some mindblowing approaches, like NVidia/Tesla using a few layers of simple convolutional neural network and static images to predict steering angles, and then some people stacked RNN on top of this CNN and made it estimate steering angle from 10 previous frames and a current frame. See e.g. selfdrivingcars.mit.edu
If you are into CV, first start with very simple static image recognition with AlexNet/VGG/Inception etc. in Keras, try to understand CNNs a bit (it's inspired by biological neurons, they can do simple things like direction detection, edge detection etc. and overlap each other's field of vision; if you look at computational photography, convolutions do something similar, so the idea is why not use a layer of multiple convolutions, then make a hierarchy of those convolutional layers, and let the optimization/learning part of Deep Learning during training figure out what exact convolutions does it need instead of force-feeding them by hand). Play with the ways to improve training (batch normalization, image augmentation etc.) Once you understand this, your mind would probably explode and then it's time to understand RNNs/LSTMs/GANs and have fun applying it on voice, natural language, generating art etc.
You'll have a blast for sure when you realize what you can now easily do! Have fun! ;-)
The acronyms aren't changing, machine learning, deep learning, and neural networks have been around for 50 years. It's only recently that code libraries like TensorFlow have abstracted away a lot of the math to the point that it's relatively accessible to normal people that can write code.
Deep learning is a subset of machine learning that utilizes more than one layer of neural networks. So these terminologies just refer to different parts of the same process. The 'process' is just tweaking a program to progressively make more accurate yes or no assumptions about a set of statistics that you give it. That's my best shot at it, hope it makes sense.
Great thing about Deep Learning is that decades of "old school" machine learning research that was way too math intensive is far inferior now. DL is actually pretty approachable and intuitive.
Treat it as non-linear optimization, then you are right in the class of most complex problems. Make a better optimizer enabling more useful applications and you can be both scientifically famous and make billions with it! ;-)
I mean, the dotcom bubble popped but websites are still here and more profitable than ever. The bubble popping doesn't mean that DL is going to go away. People will just have more reasonable expectations about what it can do.
Various products and services have shown that DL and other ML techniques are useful and profitable to implement. And corporations can see the benefits of incremental improvements. That alone will continue the momentum, even without amazing breakthroughs.
Various products and services have shown that DL and other ML techniques are useful and profitable to implement.
Very few organisations are seeing ROI on these projects. I've seen figures showing average for every million spent the return is less than half, for the quarter of them that actually get into production...
This is precisely the pattern of past winters. The technical achievements don't go anywhere. But the hype dwindles, and funding and public interest accordingly, as disappointment and skepticism grows. It doesn't permanently stop progress, just as a burst economic bubble doesn't necessarily kill an economy. Just dramatically slows it, at great cost.
Computer vision is the part of DL that is most suited to produce economic value. CV will be worth trillions of dollars in a couple of decades. All those cars, drones, agricultural equipment, medical scanners, robots and security cameras will be able to understand what they see and act intelligently. It's like the most universally useful thing since the invention of the wheel.
I said universally useful - a refrigerator is just that. A CV system can be used in hundreds of totally different applications. Like the wheel and yes, the engine. The engine is universal as well, you could see it as the upgrade of the wheel.
One of my friends is in finance, and the other in biology, and judging by the way that they talk about it, they believe that AI is about to take over the world, and they believe there is a huge monolithic black box that can solve all the world's problems. So yes, there is a huge bubble. The question is how exactly will the bubble pop? Or will it pop?
I've started paying attention to AI (again). This ML hype cycle feels like when I did optimization stuff 15 years ago. Huge interest, lots of monkey motion, then it just fell off everyone's radar.
Not equating ML to OR, directly. With today's horsepower and data sets, it is a brave new world. But we're in the elevated expectation phase of the hype cycle.
We have seen many a bubbles in the past. The AI bubble will also burst eventually. I have a first hand experience of it in the ML/NLP world, and I can safely say it is about 30% works plus 70% hype.
Large amount of data has helped, so AI systems are better than before (with lots of training data), but that's pretty much it. It is not going to replace programming jobs, leave aside solving world problems.
> It’s a phenomenon that Richard Socher, the dishevelled 30-something (or 20-something?) lecturer who just sold his company for several hundred millions yet still biked to campus, mentioned in his class: “Companies keep asking my students to drop out to work for them.”
Why would anyone not continually bike for commuting purposes once they are wealthy? Biking is such a joy whether you're 7 or 70, rich, or poor.
> Companies keep asking my students to drop out to work for them.
I've been told Berklee College of Music makes it very easy for students to leave and later come back, even years later, exactly to enable students to grab transient professional opportunities without sacrificing their education. I wonder how engineering schools stack up in comparison.
Several studies have shown that the net health benefit of bicycling to/from work is large even when you account for accidents. Not hard to believe when you hear it halves the risk of cardiovascular disease, which is the most common cause of death in the US.
E.g. these authors find the health improvement statistically increases your life expectancy by up to 14 months, while traffic accidents statistically reduce it by up to 9 days. That's a ratio of 47 to 1.
I don't think people think in terms of life expectancy. Do you? Is a 1% chance of dying on your next ride and otherwise living 100 years equivalent to not riding your bike and living 99 years to you?
Edit regarding your comment: Oh, I see why you're confused. I didn't need to do the math because I was talking about a 1-time bike ride as an example to just get the point across -- it's already scary for 1 ride. But if you want the actual math for a lifetime, there's a ~1/5000 lifetime odds of dying in 1 year of biking. That's still pretty damn high. I don't know about you but I'd rather just give up on the 78th expected year of my life and lose the 0.1% chance of dying in the next 5 years.
I'm assuming you didn't do the math there - in your hypothetical scenario I'd likely be dead within a year if I rode a bike every day. Then I obviously wouldn't do it.
But the point of bringing up the statistic is to check "is this risk worth worrying about, to the extent that I'm not gonna do thing X"? That's what a rational person does in all situations - is the risk of flying so high that I shouldn't go on holiday? No. Is the risk of falling if I climb that cellphone tower so high that it's not worth it for the view? Yes.
It sounds like we're agreeing? I was also saying, just like you, that the choice is not based on the statistics. The parent one is the one that used the statistic to justify the choice of biking.
What I don't understand is how you get a gig teaching a course at Stanford while being an undergraduate student at Stanford. Is this some sort of special seminar course or something?
I come from a similar background as the author - BS/MS in Stanford CS, but I left just as the deep learning craze was booming.
Stanford allows undergrads to propose and teach some of these more "practical" courses for upcoming technologies - a few examples are classes for NodeJS, cryptocurrency, Spark, and this one for Tensorflow. It's student-lead, very hands-on, and intended to give a specific industry experience.
The classes the author mentioned are first year pre-requisites for the AI specialization and doing research at labs, and gives enough background for a student to be instructive when explaining concepts such as perceptrons and svm's, without necessarily the mathematical rigor. This is the level needed to interact with Tensorflow.
I didn't come off of this article feeling very impressed with Stanford's CS department. Someone with contextual knowledge please explain why I am wrong.
I've had my own calculus / discrete math / math for bio courses before but that was after several years as a doctoral student and TA at Georgia Tech. I can't imagine that there isn't a PhD candidate with more experience under their belt both teaching and using TensorFlow. The author even admits they volunteered to teach the course to stimulate learning the material themselves.
I am not sure he even understands complexity of DL itself. DL can be formulated as a non-linear optimization problem; that means it's one of the most difficult computational problems and the few types of topology we know are working with current "simple" non-linear optimizers are quite miraculous; I don't think anybody understands why these simple methods work so well when we restrict/structure the number of connections between layers and why fully-connected networks have such a terrible performance even if theoretically they should be able to handle everything better. So there is IMO a plenty of space for everyone to figure out their own niche with best performing algorithm in production and enable magical things in their apps.
I get similar impression from other side.
I have met stanford phd having very poor math background. Yes it is true but unbelivable.
I think the problem is with higher education system not the hype.
Standford/cmu should make their degree more rigorous.
I am sure if cs231n include fisher vector in their course and some maths derivation in their assigment,the number of student would drop logarithmically :)
The article is an interesting mix of imposter syndrome and bubble speculation. I guess a good question is: if you know you're in a bubble and you feel like an imposter, are you right?
Sort of tangential, but I felt this exact same way when I moved from being a post-doc in neuroscience to being a data scientist. The impostor syndrome was so strong it was painful. It has subsided now a bit because I know I'm able to bring value to my company - despite the fact that I know there are much more capable and qualified data scientists (by a large margin) out there, and despite the fact that by-and-large 'ML' and 'AI' is definitely a buzzword around here. But it really, really motivates me to strengthen where I'm lacking.
The funny thing is, it took me about a year to find this position after a good amount of rejections. About a year after I got my data scientist title, I've been contacted by recruiters from places I would have never expected to be contacted from (Amazon, Microsoft, FB, etc.). Did a few interviews, and realized during those interviews that I still have a lot to learn.
For one of the interviews, they gave me a take home assignment where they literally duplicated a column in the feature matrix... I didn't catch it, and during the phone part of the interview I get asked 'do you know notice something interesting about those two feature distribution plots you have there?'
"Hrmm, no I don't. Oh, wait, they look pretty similar."
they literally duplicated a column in the feature matrix
This is called colinearity - you can check for it by comparing the rank of the matrix to its number of columns. In R qr(X)$rank. Good to add this to your EDA workflow.
I sincerely feel for the author of this post, and am not really sure how to explain my reaction while still being supportive, but...
Based on what the author is saying, part of me thinks imposter syndrome and bubble are both justifiable ways of thinking about what he's describing, but to me as an outsider the bigger problem it reveals is the way hiring and career development happens.
Without exaggerating anything about me, or without this coming from a place of jealously (although I can't deny I'm a bit jealous), it seems that I could easily teach the course they're teaching, with a deeper understanding of the material, and more justification for teaching it in many ways. I know that if I taught that course it would be fairly easy and not really stressful--fun in fact. I've taught courses on equally complex stats and math, and published in related areas.
And yet, there are no recruiters pounding on my door. If I applied for jobs most places would throw out my application for all sorts of reasons.
This person seems competent enough, so I do think there's impostor syndrome going on. Part of what they're describing is a normal process of teaching higher ed for the first time. And there probably is a bubble--the stuff they're describing is part and parcel of hype that goes along with bubbles, and although extremely useful, I think there's also a lot of problems with AI being swept under the rug.
This post really touches a nerve for me, because it gets at a problem with careers, at least in the US, which is that the bases of hiring decisions (and by hiring I mean broadly, not just as an employee) are so incredibly superficial. My guess is this person would function fine in AI, but I think anyone who knew me would have to bet that, between the two of us, I would be better qualified and better able to work in that area. But because this person taught an AI course at Stanford, they're more sought after than me, who doesn't even have a CS degree (although I do have a PhD) and certainly not a degree from an elite school.
I'm really at a difficult place in my life because I'm at a point in my career where I should be happy, and lots of people would say I'm successful, but to me I feel professionally typecast and trapped, by stereotypes and superficial appraisals. All the time you hear admonishments that degrees don't matter, etc. but then the reality is, they not only matter but matter in the most superficial ways possible, where it's not just having a degree and publishing and doing research in closely related areas, but having a degree covering exactly what is the focus of a hot plasma-magnitude bubble, from an elite university no less.
Most of the time at this point I just want a job that pays enough, and where I can live in a nice, comfortable safe place that I love. I've started to feel like the whole concept of meritocracy is a huge lie, and not because the people benefiting from it are incompetent--not because of false positives--but because of the huge problem of false negatives that lies in the shadows.
Is it really a bubble if it's producing real value?
The answer is yes. Bubbles are investment and financial entities, decoupled from the value the sector is producing, and a bubble burst can indeed destroy real value.
So AI being in a bubble says nothing about whether AI is valuable.
It's not really a bubble from the financial risk perspective if there isn't a way to "correct" or collapse it.
Machine Learning is a feature set inside an application inside a market. It's not an industry of it's own where massive swaths of an industry place their money or livelihoods, like e-commerce or derivatives.
So in that sense there isn't any bubble to burst. The majority of ML applications are happening INSIDE massive technology companies, not as stand alone companies. Even then, the stand alone companies have a product that they are selling that ML functions with. So SaaS with ML, or Image Captioning or Translation service etc...
Investing billions of dollars in companies and paying employees crazy salaries simply because their title contains the word "data", how is there no way to collapse that? One day the time comes to reap the reward and when the rewards are a lot smaller than expected funding will be withdrawn or at the very least scaled back.
Honestly? Data works. It makes companies billions of dollars. Machine learning and AI are getting increasingly good at parsing the data that already exists. It makes sense that companies, who have always made money on data, would invest in something that promises to make them even more money, and has demonstrated the ability to do just that.
aren't legacy admissions to Ivy League schools, the government protected status of Wall Street banks, generations of nepotism in Hollywood, Ticketmaster, the red-blue lock on politics, prosecution-protected city police officers, and Time Warner cable all dandy examples of sustainable rigged systems?
There is a difference between exaggeration, hype and wilful misleading and AI proponents have long crossed the line.
Now policy makers, and all sorts of busy bodies are contemplating solutions to the 'ai problem' which does not exist and will not exist for some time to come, if it comes.
Pattern matching and image recognition are valuable on their own but passing it off as AI makes a complete mockery of the word and scientific communication.
Engineers and scientists are supposed to be precise and even giving leeway for hype and excitement within the realm of what is possible.
In any fast-growing field, the distribution of expertise is pyramid shaped, with a broad base of inexperienced folks, and comparably fewer senior people.
Outsiders calling on the expertise of relatively junior people is a seems to be a pretty natural consequence of this distribution, though maybe not to the extent described in this blog post.
I've wondered if Delenn would ever show up as a girl's name. I guess now is about when you'd expect to hear about the children of people who watched Babylon 5 as teenagers. Cool!
Like other commenters have mentioned, largely due to misbranding and sensational media hype. Fearmongering from people like Elon Musk hasn't helped. But the key impact of machine learning for me is to make better and more efficient decisions that are informed by data - and that is not going to go away.