Slashdot Mirror
Machine Learning Expert Michael Jordan On the Delusions of Big Data
First time accepted submitter agent elevator writes In a wide-ranging interview at IEEE Spectrum, Michael I. Jordan skewers a bunch of sacred cows, basically saying that: The overeager adoption of big data is likely to result in catastrophes of analysis comparable to a national epidemic of collapsing bridges. Hardware designers creating chips based on the human brain are engaged in a faith-based undertaking likely to prove a fool's errand; and despite recent claims to the contrary, we are no further along with computer vision than we were with physics when Isaac Newton sat under his apple tree.
... and despite recent claims to the contrary, we are no further along with computer vision than we were with physics when Isaac Newton sat under his apple tree.
That's true, I looked into object recognition for image classification by content. Face recognition is proceeding fairly nicely but doing stuff like just programmatically classifying/tagging images by whether they contain a car, airplane, house, tree, dog, mountain .... without even trying to do things like identifying the type of airplane/dog/car is pretty much undoable in any reasonable amount of time with human level accuracy needed on garden variety PCs and tablets which is the application I'd be interested in. The fastest and most accurate image classifier/tagger is still a human. Am still looking forward to they day that changes but I'm not sure that will be within my lifetime.
Only to idiots, are orders laws.
-- Henning von Tresckow
The interview is slightly more nuanced than that. Prof. Jordan says that he can take off his academic hat and read musings on a common singularity with ordinary human awe and wonder. It is only in his work as an academic that he doesn't feel Kurzweil's ideas are relevant.
I remain sceptical of the singularity idea myself, though for different reasons. When I read Kurzweil's The Singularity is Near , I was disappointed at how in claiming a never-ending increase in the pace of technological advancement, Kurzweil never dealt with the regulatory and consumer factors, and the whole notion of how humans perceive time in general. The wheels of government can only move so fast, and so mankind's access to radical new technology outside the lab (e.g. self-driving cars, new medical tech) must slow down to match the speed of regulatory agencies. Also, consumers can be convinced to buy new shiny things, but there is still a desire to get one's money's worth out of one's purchases, and lots of people still feel their computer or smartphone from three or four years ago is still good enough. Would the market go for replacing one's tech in the shorter and shorter spans that Kurzweil envisions?
So when I read a computer scientist like Jordan admit that he sees no cause for singularity optimism within his work, I can only feel that Kurzweil's dream is a balloon being stuck with a thousand pins. Still, I continue to enjoy thinking about the subject.
The problem with computer vision is not that it's not useful, but that it's sold as a complete solution comparable to a human.
In reality, it's only used where it doesn't really matter.
OCR - mistakes are corrected by spellcheckers or humans afterwards.
Mail systems - sure, there are postcode errors, but they result in a slight delay, not a catastrophe of the system.
Structure from motion - fair enough, but it's not "accurate" and most of that kind of work isn't to do with CV as much as actual laser measurements etc.
Photo stitching - I'd be hard pushed to see this as more of a toy. It's like a photoshop filter. Sure, it's useful, but we could live without it or do it manually. Probably biggest use in mapping, where it's a time-saver and not much else. It doesn't work miracles.
Number plate recognition - well-defined formats on tuned cameras aimed at the right point, and I guarantee there are still errors. The systems I've been sold in the past claim 95% accuracy at best. Like OCR, if the number plate is read slightly wrongly, there are fallbacks before you issue a fine to someone based on the image.
Face detection is a joke in terms of accuracy. If we're talking about biometric logon, it's still a joke. If we're talking about working out if there's a face in-shot, still a joke. And, again, not put to serious use.
QR scanners - that I'll give you. But it's more to do with old barcode technology that we had 20 years ago, and a very well defined (and very error-correcting) format.
Pick-and-place rarely relies on vision only. There's much better ways of making sure something is aligned that don't come down to CV (and, again, usually involve actually measuring rather than just looking).
I'll give you medical imaging - things like MRI and microscopy are greatly enhanced with CV, and the only industry I know where a friend with a CV doctorate has been hired. Counting luminescent genes / cells is a task easily done by CV. Because, again, accuracy is not key. I can also refer you to my girlfriend who works in this field (not CV) and will show you how many times the most expensive CV-using machine in the hospital can get it catastrophically wrong and hence there's a human to double-check.
CV is, hence, a tool. Used properly, you can save a human time. That's the extent of it. Used improperly, or relied upon to do the work all by itself, it's actually not so good.
I'm sorry to attack your field of study, it's a difficult and complex area as I know myself being a mathematician that adores coding theory (i.e. I can tell you how/why a QR code works even if large portions of the image are broken, or how Voyager is able to keep communicating, despite interference on an unbelievable magnitude).
The problem is that, like AI, practical applications run into tool-time (saving a human having to do a laborious repetitive task, helping that task along, but not able to replace the human in the long run or operate entirely unsupervised). Meanwhile, the headlines are telling us that we've invented "yet-another-human-brain", which are so vastly untrue as to be truly laughable.
What you have is an expertise in image manipulation. That's all CV is. You can manipulate the image to be easier read by a computer which can extract some of the information it's after. How the machine deals with that, or how your manipulations cope with different scenarios, requires either a constrained environment (QR codes, number plates), or constant human manipulation to deal with.
Yet it's sold as something that "thinks" or "sees" (and thus interprets the image) like we do. It's not.
The CV expert I know has code in an ATM-like machine in one of the southern American counties. It recognises dollar bills, and things like that. Useful? Yes. Perfect? No. Intelligent? Far from it. From what I tell, most of the system is things like edge detection (i.e. image manipulation via a matrix, not unlike every Photoshop-compatible filter going back 20 years), derived heuristics and error-margins.
Hence, "computer vision" is really a misnomer, where "Photoshopping an image to make it easier to read" is probably closer.