Computer Vision Research: The deep “depression”

My major professor diluted the paper and added other content consistent with the previous method. Not just adding prior art to the introduction, but changing the meat of the paper so that it didn't seem like a departure.

He assured me that this would make it easier to publish, and publishing was all that mattered. There were no bonus points for publishing a novel technique, and there would certainly be extra work having to deal with referees.

I'm very glad to be out of that environment now. I noped out of academia and happily dealing with corporate B.S.

No, the reason for this is that all the easy stuff has already been discovered. 20 years ago it was still kind of easy for a single PhD student to make enormous progress in his field, but after a dozen of PhD generstions there is just not much left which can be discovered by a si gle student.

Just look at physics, they had to build a multi-billion dollar research facility below the ground to advance our knowledge. Or the satelites for discovering gravity waves, ... All of this can not be done in the classic PhD model where a single PhD student works on a topic.

tldr: After 150+ years of science there is almost nothing left which can be discovered by a single PhD student

that is good experimental data and the role of theorists here is to look how that performance achieved and why. For example, one can reasonably suspect that there is a good reason why the kernels in a well trained image recognition deep learning net do look like receptive fields of neurons in visual cortex. I'm pretty sure that there is some kind of statistical optimality in that, something similar to like normal distribution is maximum entropy distribution for a given standard variation. The same way i'd guess Gabor of neuron receptive field is something like maximum entropy on the set of all possible edges or something like this. The point here is that the great success of deep learning generates a lot of very good data for theorists to consume. You can do only so much theory without good experimental data, and in the decades before the availability of computing power (and resulting success of deep learning) there wasn't that much of the computer vision theory advances to speak about, really.

>leaves flying and various things

Newton did that for 20 years. With great success.

Well, the article is arguing deep learning has taken most of the "mindshare", the attention of most researchers. If great results are coming from other parts of the field, that would be a reason to be concerned.

It's the difference between giving a "brute force" computer proof like the four color theorem than try and come up with new theory where it's just a result from it.

But the point of the article is that there is more to computer vision. Stereo, optical flow, geometry, and physics can only be aided by deep learning so much.

Another point not mentioned is the computational power required for deep learning. Consider programming the physics for a ball rolling down an incline. You could use (1) the math itself, vs. (2) a neural net. It's clear that the direct math approach could be orders of magnitude faster than neural nets. I wouldn't be surprised if directly coding the physics would achieve 1,000x the performance of a neural net.

Deep learning undeniably works well. It works so well that it has almost completely taken over the field, sucking the oxygen out of other parts of the field. This is a little disconcerting since no one really understands how much of deep learning's success is due to a) the algorithms themselves vs. b) the massive amount of data and computational power being thrown at vision problems.

There is something to be said for some "counter-cyclical" efforts to encourage people to keep exploring approaches that aren't just slightly different network topologies, or least to ensure that they're not totally forgotten. That's the message I took away from the article.

One does science also to understand why something works (i.e. what is the theory behind it). Deep learning gives something that (perhaps) works, but doesn't give an any clue to the question why it works - on what theory does the neural net internally compute in its black box.

Based on this criterion one can even argue that deep learning papers should not be considered as science as long as the authors of the deep learning paper make no really hard attempt to present at least a partly understanding of the internal theory that the trained neuronal network executes (i.e. a very detailed interpretation of the weights with ideally falsifiable predictions). This is hard work, I know, but isn't that nearly anything in science?

It still is. Any system using structured light will have big problems when confronted with sunlight (windows, outdoors etc...).

I'm not confident there is a solution within the structured light domain to this as the beacons will (more than likely) never overcome sun intensity. We're doubling down on passive systems and reference maps.

Remember that throughout the eighties/nineties ANN classifier performance was as underwhelming as other approaches or even worse at many tasks. Now it reached a local optimum and all other approaches are being discarded.

You better separate between using lots of data (perfectly OK, in my opinion) and the number of parameters that the model uses (say, number of weights in the neural network) and where you better have good explanation for the existence of any variable and why it has this concrete weight and no other and why it is even necessary to introduce (consider any parameter that you have to introduce as some kind of physical constant - physicists invest lots of time to explain/reduce the number, so should CV researchers).