We are all essentially evil because
1) tomorrow will find some behavior that is accepted today as bad and we are all doing it
2) creating an AI that can manipulate people is too tempting for the predator in us to avoid (this is a REAL turing test - making an AI that can make tools out of humans and everything else.)
The only way to stay ahead of the 'evil' corrupting influence of new tech is to prevent it from widespread use controlled by a single entity. So, yolo is ok as long as you cannot deploy it in a cloud at scale.
So, just as nuclear weapons (the massive concentration of energy release at tremendously fast rates) are bad, so is a super AI/AGI (the massive computational ability at nanosecond scale).
No evil was ever perpetrated by institutions of learning - only business entities and governments who scaled up those discoveries caused evil.
And now for the flame bait-
So, by this argument we should elect Luddites to govern us , especially ones that are not imaginative or creative.
How is it ridiculous? The fact that purchasing a gun is the first step of shooting people is a good enough reason for most countries to ban the purchase of guns...
I don't think "because there's legitimate uses" is the differentiating factor, since that implies only kitchen knives have them. Self-defense is a legitimate use for owning a gun, for example.
There is no argument, why are people jumping to extrapolation? I'm pointing out that this DNN can be (and is being) used to identify people based on the ("any") detected faces. The possible usage is certainly related to the Parent comment issue.
Yes, and pjreddie seems to have concluded computer vision is mostly (or too often for their liking) used for the digital equivalent of those bad things.
I think prjeddie's concerns are extremely relevant. However he's not the only one working on things like this, thus it's unlikely that research and development will stop, although I certainly think such development is ethically questionable. In some ways this thing seems similar to the ethical problems facing the scientists working on the nuclear bomb. I just hope to God that this tech will be used for good rather than bad, but the way things are going with political censorship (government sponsored or otherwise) and people of opposing camps doing their best to dox political opponents–let's just say I'm not too optimistic...
> it's unlikely that research and development will stop
I know you didn't make this argument here, but I still want to point out that that's ethically irrelevant for his decision.
Or the other way around: "Someone else would have done it" is not a defense when you've built something that was clearly gonna be used for Bad Things(TM).
> Yes, and pjreddie seems to have concluded computer vision is mostly (or too often for their liking) used for the digital equivalent of those bad things.
Indeed, there are many nefarious applications of computer vision. But applications to the medical industry are plentiful too.
I see weighing up the net benefit as a tricky and a personal matter.
That's fine - that's a personal choice he is free to make. But I completely disagree with it. I also don't think that unencumbered AI research is going to lead to the overthrow of the human race by machines like Elon does.
Making cheap computer vision is just as dangerous to the "tyrant" as his supposed victims. You can already make a plausible anti-president suicide drone, A Ticket To Tranai style.
He said: "I stopped doing CV research because I saw the impact my work was having. I loved the work but the military applications and privacy concerns eventually became impossible to ignore."
Initial good results on CV doesn't mean that you realize all the ways it'll start to be used and the implications thereof.
IMO this is ethically more simple than many like to believe.
There are existing power structures (planned ones and emerged ones) in this world and the technology we create can either be used to reinforce them or to question them. Sometimes it is both and things cancel each other out and move on a sideways trajectory – but in the case of CV, it is quite clear who will benefit: those in power, those who need to quantify, control and punish the human element, but don't have the manpower (=legitimacy?) or funds (=priority?) to do so manually.
I get that working in CV is interesting and cool stuff, but the collective suffering it might help creating and keeping is something one should seriously think about as well.
Exactly. CV is interesting and can be completely harmless. But the benefits of the technology for the average person are extremely small relatively to the benefits it brings to any potential oppressive power. Mass surveillance tech can be convenient but it's a deal with the devil and I think we sometimes willfully ignore that under the guise of a perceived amorality of progress. "it's just science, it's neutral and you can also use it for good" can sound like a good argument that I usually even agree with.
But in this case it's very simple:
"Good guys" using of CV gets them things like good auto sorting in Google photos
"Bad guys" using CV can reduce the complexity of creating a fully Orwellian, big brother like surveillance state from "absurdly complex to implement and impossible to maintain" to "we can already put in place a solid implementation today and it will get better by the day".
Now I used to work in CV and as you said,I get how great and exciting the underlying tech is. But I definitely lean more towards fear than excitement these days. I also realize that it's already everywhere, with heavy research efforts and that you can't ever stop it at this point. But that really goes to show that the Yolo creator was right.
Think about it, what does the chinese people gain from CV or face recognitio right now? Maybe cool filters. The chinese government? Unimaginable levels of surveillance and control over it's entire population and it's just getting started.
Just like nuclear energy or genetic engineering, once the technology is there you can't put it back. The problem the chinese people have is not with CV, it's with their government. The way to go is having strong regulations a la GDPR, otherwise a black technology market will appear is there is enough incentive.
Also, if the state-of-the-art in these systems is in the public domain, people on the 'counter-CV' side will be able to experiment and find effective countermeasures. If all the knowledge is kept secret (i.e. it's only researched by secretive organisations) the average person will have less understanding of the technology and its limitations.
You're saying Chinese gov, but the US gov or UK gov would do the same.
I agree that society needs to come together as a whole and regulate this into law, because that's how bad actors from governments can be stopped. At least in democratic countries.
I was answering the previous comment, but yes you are also right, though I tend to think citizen's rights are more at risk in China. Anyway that cannot also be taken for granted in the US or Europe, regulation must be agreed and evolved through time.
Come to think of it, is there any tech that we would erase from history if we could.
To qualify, it needs:
- To have a negative net effect. Explosives for example don't qualify. They are certainly used as weapons and for all sorts of nasty reasons, but they are invaluable in many areas, including safety systems.
- Not to be an essential stepping stone for other, positive discoveries. For example the V2 missile made space exploration possible.
We could put all sorts of weapons in the list, with nuclear bombs in a top position. But think about it. Hydrogen bombs didn't kill anyone. Although the idea is debatable, they may even have acted as deterrents, preventing conflict. As for Hiroshima and Nagasaki, it basically ended the war, who knows for how long it would have continued otherwise, with potentially more victims than what the nuclear bombing have caused. More generally, the most technologically advanced countries are now living in an unprecedented time of peace, despite having the most advanced killing machines ever.
In the end I don't see any tech that I would put in that list because of abuse. The ones I would put there would be of the "oops, didn't know it was bad, let's stop using it" kind. Leaded gasoline comes to mind.
>As for Hiroshima and Nagasaki, it basically ended the war, who knows for how long it would have continued otherwise, with potentially more victims than what the nuclear bombing have caused. More generally, the most technologically advanced countries are now living in an unprecedented time of peace, despite having the most advanced killing machines ever.
as a german I'm always amazed how casually americans are willing to forgive themselves and even rationalize their states war crimes as good
- To have a negative net effect. Explosives for example don't qualify...
I was more idealistic about this kind of thing at one time, but it really is easy to rationalize the "Work on what you want to, don't worry about the end use" attitude that I have now. All I needed to convince myself was a good example.
Imagine that it's the early 1980s. Reagan is in office, the Cold War with the Soviet bloc is still very much a thing, and Star Wars is ramping up. Every other week it seems that somebody proposes yet another batshit-crazy weapons system. You're an engineer with progressive political views, and your bosses are asking you to work on a vast, global satellite network that will allow the military to locate both targets and assets with pinpoint accuracy anywhere on Earth. As far as you're concerned, ol' Ronny Raygun can fuck right off, and you tell them as much. "I'm not working on anything like that!"
20 years later, it turns out you missed your chance to get in on the ground floor of the most important public utility since the telephone system, all because you could only see the destructive uses for the technology.
For me that's hypothetical since I was nowhere near old enough to be employed at the time, but it's easy to say the same thing about applications like UAVs, autonomous vehicles, and ML/AI in general. CV is nowhere near enough of a defense-centric technology to justify refusing to work on it, IMO. Someone who refuses to work on CV on ethical grounds is walking away from their share of our technological future, just like the hypothetical engineer who refused to work on GPS.
We could put all sorts of weapons in the list, with nuclear bombs in a top position. But think about it. Hydrogen bombs didn't kill anyone. Although the idea is debatable, they may even have acted as deterrents, preventing conflict. As for Hiroshima and Nagasaki, it basically ended the war, who knows for how long it would have continued otherwise, with potentially more victims than what the nuclear bombing have caused. More generally, the most technologically advanced countries are now living in an unprecedented time of peace, despite having the most advanced killing machines ever.
I tend to agree with the general notion that nukes are a net win for world peace, but it's debatable whether the net death toll due to warfare has been that much lower in the post-Hiroshima age. Superpowers just conduct proxy wars nowadays instead of beating up on each other in person. If you added up the civilian toll of those proxy wars, it would probably be right up there with many WWIII scenarios, but since those conflicts are happening somewhere else besides major American or Soviet cities, nobody much cares.
The other concern I have is a relatively new one: people are going to forget what those things are and what they do. Eventually, the last person to see a nuclear explosion in person will die of old age. Long before that happens, morons with microphones will deny that Hiroshima and Nagasaki ever happened, just like they do now for events ranging from the Apollo landings to Sandy Hook. Others will take the position that nukes are just bigger versions of regular bombs, nothing that special.
So long term, who knows... maybe it would be better to put that genie back in the bottle if we could. Dunno, and in any case, it's hardly the same thing as CV. I can understand if someone is reluctant to work on nuclear weapons technology, but that understanding stops well short of refusing to work on CV.
>it is quite clear who will benefit: those in power
I think it's more likely that you want to detect a person so you don't hit them with a car than trying to hit them with a drone missile. Detecting cancer with CV, and other improvements to diagnostics also save lives. If you could be working on these technologies and stop, your decision could cost lives.
The same neural network can be trained for both problems. Advances in medical image neural networks translate to advances in people detecting neural networks.
Of course, they are not directly used without changing the training data, etc. However, if we look at the big picture to make a general statement, then the advances brought by Computer Vision on natural images often help push research on medical images forward, albeit with a year (or two) of delay. One such example is U-Net (https://en.wikipedia.org/wiki/U-Net).
>but in the case of CV, it is quite clear who will benefit: those in power, those who need to quantify, control and punish the human element, but don't have the manpower (=legitimacy?) or funds (=priority?) to do so manually.
Technology that exists for surveillance can also be turned on the surveillants. The most relevant case probably being police abuse being caught on smartphone cameras. These tools don't just discipline citizens, they also discipline the police. If I'm in a room with someone in a position of authority far above me, I'd rather have the camera on both of us than none of us.
So it's not actually that simple, and I don't see opting out as realistic or helpful, because other benefits these technologies bring, for example security, will always convince the population to drive adoption forward.
I work in FR, at a leading company. I do not write any ML code, I'm an applications developer with a background in visual effects.
My attitude is being there in the developer group influencing how the applications behave, interfacing with our management, sales and clients and being a voice unafraid to raise ethics questions within these groups is my method of knowing what is the state of this dangerous technology. I'd rather be there influencing its development and use than on the outside, frankly blind to what it's in-deployment capabilities are.
You can only influence these things so far. The company ultimately needs to make money and sometimes the money is coming from people that want drone-guiding or surveillance technology and so if your suggestions are leading too far away from that, you'll be fired.
I mostly agree on your thinking but I think it is not so simple to decide.
If we focus on reinforce vs question part. It is kind of prisoners dilemma. If we assume governments(reinforcers) will anyway work on this technology, and lets assume they will end up with some lower quality version (lets say 50/100) . Questioners have 2 options:
- dont work on this, accept 50/0
- or work and improve both sides, 70/20
Without deeper context hard to decide which will be better for questioners.
Would you prefer to have a gun against rifle, or no weapon vs gun?
The question is, if we are so aware of the dangers posed by these power structures, agreeing that the technology itself is "apolitical", then why have we allowed power to be concentrated so highly in a given institution or individual? It seems that, if we are concerned with ethics, we should be pushing much more for the dilution of power back to the people, rather than "taking sides" in terms of who to work with or what to work on.
Just trying to provoke an answer with a little humour ...
It is obvious to anyone that one of the primary uses of computer vision is to watch other humans at scale. This cannot be surprising to anyone in the field yet now there are ethical concerns mixed with politics through the roof.
Maybe we can be more precise about this?
The work here is a technical achievement but there are some weird comments here which I think has something to do with the author being Chinese.
Focusing too much on negative aspect of thing will lead you nowhere.
Manhattan project gave >500 research papers. Gave us Iodine 131 and other radio nucleotide which we use in medicine. And gave us lasting peace. So was it bad or not?
Is gene editing bad? was the internet bad? Was the dude who invented round wheel bad?
Yes, using atomic bombs was bad. As others say it is debatable if it helped end the (already won?) war. And in any case, _lasting_ _piece_ where? Are middle east and Africa not part of this world?
I find your comment is entirely missing the point and low effort. Asking whether random techniques, inventions, inventors were "bad" or not makes us much sense as asking:
Is the sun bad? Were dinosaurs bad? Are the aliens bad? Is life bad?
I believe that the USA and USSR/Russia have been at war with some other country directly or by proxy more or less always since the end of WW2. Yet we had no other major world war because of the mutual assured destruction. That kind of peace might not last forever but it was an unusually long period. A wonderful change for the best if you lived in Europe, no change at all if you lived in one of the countries that took the turn to be the target of one or both the powers. Add regional powers operating on their own will now.
one human brain can only perform some 10-12 hrs of facial recognition work before needing to take a break, and also do so relatively low speeds.
One computer brain can be copied for free, and deployed to thousands of computer clusters and work 24/7 on facial recognition, at a fraction of the cost.
I'm looking for something to run on a Raspberry Pi, to detect humans on a security camera. The built-in camera software has false triggering, esp. on windy days.
When looking at these projects, how do I figure out what hardware they're aimed at? This one mentions NVidia/CUDA.
Is there any sort of hardware abstraction layer that YOLO or R-CNNs can operate on? Can I use any of this code (or models) for my R-Pi?
'round about 2004 I built a pentium based 'motion' recorder. It kept a circular buffer of images that were spooled to the output stream when motion was detected. Motion was determined by optical flow iirc - the OpenCV call returned an array of blob center points, size, and velocity vector. If the blob was large enough and the velocity vector made sense (eg, horizontal as in walking or driving at an appropriate magnitude) it was considered motion. Reduced leaf flutter, branch waving false positives to effectively zero. No ML required.
ML is too liberally applied without understanding how or why it has triggered.
I've lost track of the original quote but the spirit of it is: "It's artificial intelligence while we don't understand it. Once we understand it, it's computer science"
I use the Movidius NCS and it's pretty much the optimal solution for light OpenCV work. It draws very little power, but as the Pi itself is powered by USB, don't expect to run much else off of USB on it.
I use it in the exact same scenario, a Raspberry Pi (Zero W) with a camera with motion detection and notifications on movement, my implementation may be specific though.
Each of my Raspberry Pi cameras runs motion (https://motion-project.github.io/index.html), and recorded files are stored on a NFS share. Each camera has it's own directory within this share (or rather each camera has it's own share within a parent directory), and the server then runs a python script that monitors for changed/added files, and runs object detection on the newly created/changed files.
If a person is detected in the file, it then proceeds to create a "screenshot" of the frame with the most/largest bounding box, and sends a notification through Pushover.net including the screenshot with bounding box.
There implementation is not quite as simple as described here, i.e. i use a "notification service" listening on MQTT for sending pushover notifications, but the gist of it is described above.
Edit: I should probably clarify that my cameras are based on Raspberry Pi Zero W. They have enough power to run motion at 720p - at around 30fps. Not great, but good enough for most applications. I've since migrated most to Unifi Protect instead. A little higher hardware cost, a lot better quality :)
I played with the Movidius stick for a bit but it didn't seem like Intel was committed to Movidius tech in the long term. So I'm curious what the "next" thing will be.
They were/are committed enough to put out a second version, and I guess these things are still pretty much a niche market, I mean they specifically target embedded applications.
I think the next thing, which is already happening, is embedded “AI” processors on SOCs. Apple is doing it with their A12 processors, and it’s probably only a question of time before you can purchase “general purpose” ARM chips with “AI” capabilities.
I love how every new YOLO project inevitably leads to the discussion of the ethics. At the very least more people will be wondering if they should also be taking ethics into consideration wrt their lines of work.
Pjreddie is a giant for this. It is a real contribution.
Trying to think of some applications for this. For example one could create a mechanism that watched people entering and exiting a shop providing the shop owner more quantitative data that he could use to optimize his sales.
Or you could have it watch a soccer game. Generating all sorts of data on how the game went.
Entering/exiting buses for automatic passenger counters is more important than ever now. Being able to broadcast GTFS-Occupancy in real-time when only 50% (or less) of the bus can be filled with passengers, is a real issue transit is facing today.
"Bflops"? I'm guessing this is a measure of the total processing power needed, in billions of floating-point ops, and not a measure of operations per second?
Thank you for the link. I'm still curious about the possibility of taking a detection algorithm/network and just running it in reverse. Is that feasible? Are there people doing it?
I have used Digikam because I thought it was 4 years newer but it was a complete and utter disaster. Huge problems (slow) when using it with a lot of pictures, face detection all over the place and a shitty interface to tag people.
Picasa on the other hand has no problems with huge libraries and the tagging interface is actually fun to use.
Downside to Picasa is that the initial indexing is very slow, as it only uses 1 core.
Also (both Digikam and Picasa do this) software needs to write metadata into the JPG file. No sidecar allowed.
Face detection is a critical and difficult part of face classification, it's not like we're decrying the invention of the transistor. It's more like working for a bullet manufacturer and hoping they'll be used for hunting deer.
It ends wherever you set your ethical boundaries, and it's certainly not black and white, but I'd say the man who builds fireworks from gun powder has an easier time than the one who puts it in steel casings.
Xolo does feature vector extraction on face regions. Extracting those regions from larger images is something else’s job.
Zolo does feature vector similarity analysis on facial feature vectors. Extracting those vectors is something else’s job.
Oh look, now every authoritarian government has free access to never before seen levels of data harvesting. But nobody has to feel any guilt because they only contributed a third of the machine. Hooray!
If open source models aren't developed, and even if they are, governments are going to develop their own. At least this way everyone else is on equal footing and not at a strict disadvantage. This tech could be used to identify police when they are in riot gear and cover their identification, for instance.
Well, you often can see police officers' faces through their visors, just as you can sometimes recognize people through fabric masks. Police don't wear gas masks all the time, only when they're about to or immediately after deploying gas.
The police I saw at Seattle protests did indeed wear de-identification equipment even when not actively deploying chemical weapons against the civilians. I have photos.
The future is coming, and the tech that can enable authoritarian behavior will come no matter what we do as they're just tools. It's been here and around us in ever-increasing forms for decades and yet we haven't necessarily devolved into a big brother state.
What we should be worrying is actual usages of grand scale citizen-control and monitoring projects that are enabled by technology. Think China, not UK or US.
We have to worry about it in the US and the UK too. I mean, monitoring is already widespread in the US, UK, Canada, etc. Filtering, and therefore control, was almost reality in the UK last year, albeit limited to pornography. But it's a slippery slope in both cases.
Imagine an app like Pokémon VR but instead of virtual pocket monsters it targets members of <outgroup> whose faces are detected by cadre smart phones, then tracked, flash mobbed and dealt with, for the crime of making members of <ingroup> feel unsafe.
Watching leadership supine and carefully uncritical of burning, looting mobs offers little confidence that they will stand in the way of this. After all only <outgroup epithet>s have anything to fear.
The submission is an implementation of a core task in computer vision. Your response is an appeal to vividness unrelated to the submission except in its recruitment of computer vision to sell the FUD.
If researchers getting good at something is sufficient priming to cause you to direct your imagination toward hyperbolically negative outcomes, the problem on your hands is a constitutional resistance to further progress in the research area.
In that case, challenging readers to produce arguments on the finer details of the narrative you've painted in support of the technopessimism is bad faith rhetoric.
I'm similarly unclear on what sort of response you expect.
There weren't many comments when I replied.
My criticism was tailored to a fairly specific phenomenon: asymmetrically imaginative doomsaying that appeals to a vivid vignette / sketch of an adjacent possible future featuring some hyperbolically elaborated extension of trending tech, like Flash Mob Gone Wrong[1] and Slaughterbots[2].
The only way to stay ahead of the 'evil' corrupting influence of new tech is to prevent it from widespread use controlled by a single entity. So, yolo is ok as long as you cannot deploy it in a cloud at scale.
So, just as nuclear weapons (the massive concentration of energy release at tremendously fast rates) are bad, so is a super AI/AGI (the massive computational ability at nanosecond scale).
No evil was ever perpetrated by institutions of learning - only business entities and governments who scaled up those discoveries caused evil.
And now for the flame bait- So, by this argument we should elect Luddites to govern us , especially ones that are not imaginative or creative.