A New Batch of Deep Learning Startups to Watch for in 2017
scrolib.com
scrolib.com
Deep learning techniques are indeed very efficient at solving a very large spectrum of problems, but it does not mean that every company using DL is remarkable. That's remarkable is the goal they achieve or aim to achieve.
If they use decades old technologies in a meaningful way for their target industry, does it make their company less likely to suceed? No it does not, for all that matters is solving the customers' issues...
My company use DL in several fields, but what we sell is simply a solution to customers needs...
The fact of the matter is, if you are smart enough, you can get acquihired for a lot of money. DeepMind, Geometric Intelligence, Deep Blue Labs, Vision Factory, TupleJump, Turi, etc. are all examples.
They happen to rely (partly) on DL to deliver this value, but they got acqui-hired not because "we are doing deep learning" but because "we have technology that can lower your costs/increase your profits"...
They also provided value before but nothing has been made public about it yet, unfortunately.
- Applied deep learning is still not easy. Not many people have know-how about training models. There are no easy answers in hardware space.
- Translate academia success to businesses. After DL achieved remarkable things in Academia, people expect that same success will be translated to business. Business success mostly only came from Google.
I built a company on translating academia research to business so I can totally relate to that feeling. I consider that a company exists to provide value to its consumers. If a simple linear regression provides value, why go the DL route?
I am absolutely not saying that DL has no use-case, we use it internally and on customers projects, I am just arguing that if decades old stuff does the job, don't bother with DL, in particular because applied DL is indeed very hard.
To me, it is like an extension of the "if it aint broke, don't fix it" philosophy: "if it can be fixed with a simple solution, dont over-engineer".
I often hear young software engineers laugh about banks still using COBOL for their core business. COBOL is lame, it's monolothic, etc. It may be all fashioned (and the little COBOL I've done I've hated it :-) ) but it goddamn works, why touch it and risk breaking stuff?
If a new technology efficiently solves a use-case then fine, go for it, but I hate the "hype" phenomenon. For instance, about 5 years, the hype was RoR, 2 years ago Scala, now it's DL and Pony, etc... I do some pretty nice DL, it's useful. I really like the concepts of Pony. But I'm not going to promote a deep learning framework in Pony to replace SWIFT money transfer, even though I'd be ticking many boxes of hype with that...
Anyway, I am working on deep learning and it enable us to solve problems of our users better than previous computer vision approaches.
Many computer users run some form of mathematical optimization everyday, without realizing it. Through a peculiar turn of events, the optimizations which are widely used today are often called Machine Learning, and many of its users are not aware that it is basically non-convex optimization, or high dimensional function approximation.
There really is Deep Learning, and these people are using it, but it is just a part of the optimization mathematics they use. Deep Learning is the kernel: the program in the system that allocates the machine's resources to the other programs that you run. The kernel is an essential part of an operating system, but useless by itself; it can only function in the context of a complete rigorous mathematical framework. Deep Learning is normally used in combination with other ideas in classical mathematical optimization: the whole system is basically stochastic gradient descent with GPUs added, or large-scale numerical optimization. All the so-called Deep Learning algorithms are really algorithms of generalized function approximation!
Deep learning practitioners are aware that they're using non-convex optimization, stochastic gradient descent, and so on. In fact, those topics are core to modern deep learning research and explicitly acknowledged in most published research: LSTMs were invented to solve the vanishing gradient problem, Restricted Boltzmann Machines were used as a pre-training step to avoid local minima, and optimizers like ADAM have explicit guarantees about things like convergence.
You may know all this stuff already--not sure, based on your comment above.
If not, here are a few example papers from mainstream AI researchers explicitly talking about deep learning as an optimization problem or function approximator: [1] Why Does Unsupervised Pre-training Help Deep Learning?http://www.jmlr.org/papers/volume11/erhan10a/erhan10a.pdf [2] Multilayer feedforward networks are universal approximators: http://www.sciencedirect.com/science/article/pii/08936080899... [3] Deep Learning, Nature. http://www.cs.toronto.edu/~hinton/absps/NatureDeepReview.pdf [4] Adam: A Method for Stochastic Optimization. https://arxiv.org/abs/1412.6980
It really depends. It's now perfectly possible to use deep learning as a black box, and I've already seen many people use it that way.
Just like you don't have to know how a web-browser works to be a web-programmer, you don't have to know how deep learning works to (e.g.) train a classifier.