That's the big plus of AI algorithms. For instance, all voice recognition algorithms use a patented algorithm. Nuance holds the patent.
But, the reasoning goes, because this was learned, and there is no code in there implementing that algorithm (just "weights" implementing an unrolled version), that code does not violate patents.
It's not a bug, it's a feature. Know any valuable algorithms ? Figure out how to learn them.
Whether it's a big plus or a big minus depends on the application. I probably wouldn't mind if voice recognition software in a device like Google Home or Amazon Echo uses an opaque and unverifiable algorithm. But the creator of an algorithm that controls something that can kill me - like a car or a plane or a nuclear power plant or a robotic surgery tool - had better be able to prove that their software works safely.
But if the voice recongition is part of your driving experience and it bugged out you can end up in a car crash. One such example is Siri and Google Home must hear the phrase to starts with 'Siri' or 'Google', but what if your car's has a bug and it heard you say 'stop the car'? Or that came out of your radio host's mouth? I remember a couple years ago some news covered this, you can trick Siri to think it was you commanding if Siri hears a conversation starts with the phrase 'Siri'.
The problems you mention are valid (bugs that result in accidents & closed source algorithms) but the example is unlikely. Surely, carmakers will not allow vehicles to respond to commands in such a literal way that would definitely result in an accident. For example, if the voice recognition hears "stop the car" the response will be to stop the car, but safely rather than immediately/unsafe. This goes for all other valid commands. The alternative to a safe execution will be a prompt requesting clarification because the command cannot be safely executed, cannot be understood, etc.
That's a bit like cutting off your nose to spite your face.
Sure, you can get around some silly patents, but you have only solved that problem (and an implementation at best): because you can't get the algorithm to explain itself, you don't know why it works, or how and you hence can't use that knowledge to build something better or increase your understanding.
That's another big plus of machine learning algorithms. You don't know how they work ! You don't need to !
I don't know anything about how to tell syllables from eachother. I don't know how to transform images so that I can transform an image of a 3 into a 3. I don't know ...
And yet I can make programs that do all those things ! That's great. That's the big advantage here.
This article makes the assumption that we are learning a complete model that goes from sensor inputs to control outputs, but I don't think anyone is doing this outside academia. There's a whole lot less controversy when we use deep learning to do scene understanding, where we understand at a high level the model is recognizing entities in its sensors, and we can evaluate whether that subsystem failed, etc.
I've always heard that argument in favor of decision trees or random forests, yet those decision trees had 400k nodes :). So no one ever really looked at them, but in theory you can could check the long node paths doing arbitrary splits on weird features :).
Apart from that, the strength of DNNs is exactly that complex decision making compared to, say, the simple algorithms physicians learn and manually apply for diagnosis. Those are obviously vastly underfitting in many cases.
If we want to "understand" what a network does, that really means we want to disentangle cause and effect and spit out simple algebraic models for it after distilling them from a training set.
To the extent this is even possible - which is debatable, for all kinds of reasons - we're going to need a different set of tools. ML is not the right tool for that problem.
Something similar to ML may be, but ML itself definitely isn't.
This is kind of a silly strawman in some ways, simply because all software - including the code helping fly jumbo jets, steer oil tankers or run MRI machines - is written by fallible humans, and is generally considered safe only because of QA testing, rather than code analysis. There are some rare instances of insanely complex code having every line thoroughly vetted like those in NASA projects, but pretty much everything else out there is simply "good enough" until a flaw is (inevitability) found and fixed. The decision trees generated by AI will be no different. Until, I guess, an AI can perform the analysis of the code of another AI... cue Inception music.
the jews will destroy everything and demand that everything of genuine value be debased until it is worthless.
even the concept of clear thinking is not safe from the jews. they will lie and cheat if you let them, and they will do it until you hurt them enough so they can no longer do it
the jews should be hurt so much they will never try to speak again, they will be dumb robots who never speak
'Getting a car to drive this way was an impressive feat. But it’s also a bit unsettling, since it isn’t completely clear how the car makes its decisions.'.replace(/car/g, 'human')
17 comments
[ 2.7 ms ] story [ 47.9 ms ] threadBut, the reasoning goes, because this was learned, and there is no code in there implementing that algorithm (just "weights" implementing an unrolled version), that code does not violate patents.
It's not a bug, it's a feature. Know any valuable algorithms ? Figure out how to learn them.
Me: "That's funny."
Friend: "Do you want to get lunch tomorrow?"
*car assistant: Do you wish to stop the car?"
Me: "Yeah." (responding to my friend's question)
This is a valid response. Yes, but this is the case if the assistant has a bug, so the safety meter isn't working quite properly.
Sure, you can get around some silly patents, but you have only solved that problem (and an implementation at best): because you can't get the algorithm to explain itself, you don't know why it works, or how and you hence can't use that knowledge to build something better or increase your understanding.
I don't know anything about how to tell syllables from eachother. I don't know how to transform images so that I can transform an image of a 3 into a 3. I don't know ...
And yet I can make programs that do all those things ! That's great. That's the big advantage here.
Apart from that, the strength of DNNs is exactly that complex decision making compared to, say, the simple algorithms physicians learn and manually apply for diagnosis. Those are obviously vastly underfitting in many cases.
To the extent this is even possible - which is debatable, for all kinds of reasons - we're going to need a different set of tools. ML is not the right tool for that problem.
Something similar to ML may be, but ML itself definitely isn't.
even the concept of clear thinking is not safe from the jews. they will lie and cheat if you let them, and they will do it until you hurt them enough so they can no longer do it
the jews should be hurt so much they will never try to speak again, they will be dumb robots who never speak