I think we have to start thinking about a way to make it work. There is pressure in some areas for papers backed by things like Jupyter notebooks, and machines make accessing the horde of information associated with a paper easier. We also need to disintermediate Elsevier and other bad actors, they're really causing a lot of trouble at this point.
The problem IMO is that the internet is too much of a bazaar, and too little of a library. If there was a means to mark off the 'library' content of the internet such that if you asked for a resource at time t, you would get that resource at time t, many of these issues would go away. Sort of the wayback machine on steroids. Yes it would be hard, but not NP hard.
This could also allow the use of blogs, in much the same way as private communications can be cited in papers. For example, "We elected not to run fluorine experiments due to advice from this source -- http://blogs.sciencemag.org/pipeline/archives/2010/02/23/thi...
I know it isn't easy, but we really do need to embrace the internet with respect to scientific literature, not keep walling it off. Cornell's arxiv seems a good start.
We all know the systems are biased, because we feed them biased information - Cathy ONeil did a great job on this with "Weapons of Math Destruction"
What puzzles me is why he doesn't just provide the counterexample to scare the hell out of everyone - if he inverts the racial distributions in the training datasets, then you'd have a system where the shoe was on the other foot, and both argument and fuel for removing bias from the training sets - the recognition systems only know what they know, and they don't really understand that they might not know, they only know the confidence they have in knowing - hence this is my favorite happy place example of the problem --https://youtu.be/UFVB5rnqjyY
Enhanced radiation weapons, aka neutron bombs - there was initial excitement followed by a mass exodus - the govt held on for a while longer, but they move slowly - the labs and contractors visibly got cold feet over this
I'm wondering if it might not be a good solution for >some< layers in a deep network, i.e. it's just another building block - practically I'm wondering just how far this puppy can distribute over gpus, and if you could pipeline the whole thing - i dunno tho, as far as I've gotten is that it might be an interesting experiment - and I don't want to get completely sidetracked by an ODE solver :-)
I've long wondered about this - I see and understand the comments by people pointing out this paper may a bit shallow - in the same sort of vein, I have this on my desk to grind thru - interested if anyone else has looked at it -- https://arxiv.org/abs/1806.07366
>"Given that society created the environment for these tech companies (and their employees) to acquire their skills and thrive, does not society have some moral/ethical lien on how those skills are applied?"
Yes. And that's the rub. I had the good fortune to work in research for the military for almost 10 years. They have more patience and willingness to adapt than any commercial company I've known. Their time horizon spans a much longer interval than most corporations. Did neural networks out of the PDP groups work in the '80s, and when things were too slow, they went yup and bought a bunch of Mercury computer cards to speed stuff up without squawking or delaying.
Now I do work in stereoscopic machine vision. I'm pretty certain that biological or machine, the lowest levels of the visual stack are just targeting systems based on feedback loops. Here I agree with Dr. Russell at Bezerkely. We must not, can not, make autonomous killing machines because the consequences are potentially too catastrophic. So I refuse to work on anything resembling autonomous mobile weapons systems, as do many many others in this field.
This has happened before - we were playing with enhanced radiation weapons in the '70s and '80s (neutron bomb), and over time pretty much every participant decided it was a very bad horrible no good idea. I would argue that for talented players capable of dragging the future into the present, the issue is not that the military is bad, it's whether a specific application they want is unequivocally bad. It's not about who, it's about what.
Especially the cartoon captioned something like "2+2=3.9erstėsdgdfg - the system is 99% working" - both funny and accurate, after a fashion :-) Too bad all we got for the MVC book was "The black art of MVC programming" :-(
I agree with all but one of your points, the projection of her mindset. I've known many good technologists driving for the next big thing, hell I've done it more than once. Usually we're a bundle of quivering insecurities with an unhealthy focus on all the things that are failing in the approach and a level of fear approaching utter terror that the entire thing is going to end ignominiously. My take is that she may have had a vision, but no real understanding of the process to get there, hence 'she' never would have made the tech work and would have known that she could not anchor any of her claims to reality. That's pretty much deliberate fraud. The other alternative I see is she was pretty much brainwashed by her partner, which might be some defense.
LIDAR is certainly useful in building up point clouds, especially when coupled with RADAR, however the devil always lies in the details
LIDAR is higher resolution than radar, but often slower - counterintuitive to be sure, but such is the case - i.e. this large volume of data can lag reality by some small delta, and at 60MPH, that can add up.
Uber does use LIDAR and RADAR, however the informed external commentary I've seen, admittedly guesswork based on the NTSB reports seems to indicate it was a fusion error behind the AZ fatality. Sensor fusion is a beastly complex problem on top of miserable calibration exercises :-(
More importantly, AI only exists in the minds of marketers, media, and the improperly informed. This is all just pattern recognition - we're a ways off from having a system understand that if an occluding obstacle moves out of the way, then high priority previous unknown information now exposed needs to be checked. I'm sure such tests will get hardwired into current systems, however the system is then limited by what does get hardwired into it.
At the end of the day a visual system, no matter how sophisticated is still only a visual system.
No - as the restrictions fall off, the probable number of sellers rise and the price falls - some of the current value of his holdings reflects the fact that they are not at play in the market - its the old supply/demand at work
AsI recall, Terry Winograd later claimed SHRDLU was the beginning of the end of the strong AI trajectory, at least at that time. I believe he, or someone close, said the central weakness of the system was that when it failed, it did not have any understanding that it had failed, much less why. Seems this is the current SOTA in chatbots now, given the somewhat snarky examples in the article.
I just love hacker news. One of the last bastions of people who think on the internet. This comment is succinct, dead on the money, and absolutely terrifying. My hat is off to you, I had never managed to formulate the risk so concisely. :-) Keep on doing this, and you help keep us all honest.
Not dramatic - the robots are not expected to reload - one robot one kill - simple as that- mass produced plastic, lots of robots, no intention whatsoever for reusability - the risk is that this means you dump thousands to millions of these stupid little drones, many make mistakes, and many others take naps and then wake up and do more damage when everyone thought it was all over. These are disposable killing machines, not Reapers. Furthermore, at todays prices, a cloud infrastructure and a million dollars can net you 10K lethal drones. There's no finesse, just brute force.
MAD was based on the concept of the 'Nuclear Club' - the barrier to entry was the mass defect ratio, which you'd have to calculate on your own, and then you'd confront nation-state level expenses of turning that knowledge into practical engineering. It had an inherent cap on the number of possible players. Not saying it was nice, it was just manageable. The barriers to entry here are much lower. For smaller players, lethal drone prices would be at the several hundred dollar range with a six digit investment in infrastructure, and I'd have full confidence in American defense contractors to drive that price much lower, in order to protect their margins, and at scale, they'll mess up their inventory control. The only choice is to recognize as a species this cannot be done, or it will be done. This equation is not about the few agreeing not to kill the many, it is about the many realizing that this is a road to an extinction level event. I don't mean to be dramatic, but I have a system I could make kill if I wanted to, and I paid for the whole thing on credit cards. Its too easy and too efficient. The number of players can't be counted on one hand, it's a tech lone individuals could deploy. If they do then others will, you net an exponential growth rate in deployment, and then the game is over. At the end of the day, all of the pieces are already here - a motivated individual can do a lot of damage, several motivated individuals or groups can get into a squabble and the damage rises exponentially. It is simply not an avenue we can accept. The limit to entry is not financial, it is not technical, it only can be that we must not. I'm pleased Google has defined their ethics as a company, I just argue in this case we need to define our ethics as a species. We cannot build these because if we do they will get loose, they will be used, and it will be a tragedy. We must all agree that arming autonomous machines is beyond the pale, for any one, for any reason. As soon as one is built, your destruction is guaranteed, regardless of whether you too build one. .
I've done work in this space, and my stereo cameras are about to go airborne on Air Swimmers (Bruce the Shark :-) ) in my house. In the movie in the linked document its just looking around my office and showing parallax error - I have an app that lets it track, and I spent an afternoon trying to make it loose me - the only way I won was because I jumped from where I was, crashed into the toolchest, and fell to the floor - it actually tracked me successfully, what defeated it was the fact that I was using cheap servos and it snapped an internal axle on the cheap yaw servo when I hit the toolchest. We are helpless, there is no way around it, and there is absolutely nothing we can do about it except realize our survival lies in not going down this road. https://www.linkedin.com/pulse/low-cost-high-volume-stereosc...
Yes. I can only refuse to play and hope the Russians et al come to their senses, the United Nations starts to work better, etc. If we go down that road, it will be the end of most of us, if not all of us. We really have no chance against a machine that's decided to shoot us. As a species we need to have zero tolerance for it - we won't be rendered extinct by snazzy terminators, we'll be killed by the 21st century version of a Victorian steam loom. Its truly a horrible problem, because the solution can't be creating autonomous systems to kill other autonomous systems. The rest of the distinction is just data. The only way to win is not to play.
What about street level city/town surveillance? Of course all the cop shops want to feed all their wants/warrants into it. I have no problem with this if there is a legal framework that supports it. More importantly, there are toddlers that get loose from their mothers because they are fundamentally greased pigs, or there are older members of our society who may be prone to wandering off in a daze - the ability to locate them within a very few minutes would be a good thing indeed. My own take is that I will look, I will try and recognize, but I will NEVER target. Machine is too good, sacks of meat like us have no prayer. That said, I've talked to a couple of cop shops, my position is that I will work with them when there is a framework that defines a legal basis for that level of tracking of people who clearly wish not to be tracked, even if they are bad actors. It's not a question of how far away you are, its a question of what you will track, why you will track, and what structure our society has put in place to get the benefits, not the risks. Looping back to the main point of the post, looking is OK, tracking requires societal governance, and any autonomous lethal capability, never mind action, should be absolutely forbidden, not just by law, but because we don't all want to die.
True - I often make the heretical argument that SL vs RL is just a question of where the labeling comes from :-) You are correct that the tooling is weaker in this space, but it is growing - my point is only that there's a difference between knowing how to use the tool and knowing how to make the tool - making the tool can liquify your brain -- using someone else's tool (assuming its a good tool) will simply give you headaches from time to time :-D
That is going to be a tough one, and possibly even impossible based on where current ML tech is leading. A laudable goal, but taking a complex system with a training time of many GPU years and asking it how it came up with the answer basically nets a very large pile of numbers (weights) tied together in a complex multidimensional relationship that we just plain can't follow outside of the system. Right now the practical focus is on trying to stop feeding the systems biased data. Your example is spot on, save for the 'not able to talk to someone' which is just googz being too aloof and too cheap
That's a pity - as someone in stereo machine vision, metrology is something we really really need for a huge range of applications - that said, if its feeding an autonomous armed targeting system, I agree. We really need to decide as a species we want to continue being a species and realize that solving problems with each other by force is coming to an end. I'm not speaking from a pacifist viewpoint, just a practical one - there's pretty much no way we're not going to continue evolving vision systems in all their hyperspectral glory, and they can do a lot of good - we just cannot arm them. Ever. And surveillance is a complex question - there are enough positives it can offer not to rule it out a priori
We only call it AI - creating an autonomous ML system that is designed to kill anyone can lead to us all being killed by math and plastic - it's a singularly bad idea, so much so there is no scoring mechanism that justifies it - too much chance we end up with this -- https://youtu.be/9CO6M2HsoIA
It depends on whether you are implementing a library to support RL or using one that does - in the former case, yah, you better have a grip, in the latter, you need to understand enough to know how to utilize the library and not feed it nasty things.... this isn't to say you can be mathematically illiterate, however there are levels to required knowledge - the same as you don't need to know how to make a compiler to use one
It's a cultural problem that has arisen across much of tech and is embedded within it - if you're going to have a science based business, then science says that while you may have picked your initial hypothesis because you love and believe in it, the only thing that matters after that first step is measurable and provable results. I see this a lot in the machine learning space "We've got this great ML solution all ready to go.... annnd we don't even know what linear regression is" Start your journey on a dream, but navigate by facts, you must.
Depends on your model - in the majority of cases training is a separate act from using the model, and just because you trained it doesn't mean it's well fitted. You end up caring about training speed because you don't know how many training runs you may have to do, grumpily fiddling with the model and hyperparameters after each training session trying to get a good model. Technically possible but probably not practical