A person's history has to be stored somewhere. A devoted Sean Hannity fan is somehow distinguishable from a Rachel Maddow devotee. In the end, it's all 1's and 0's. I believe I would find it satisfying.
That said, I'm open to being disabused should a piece of reasonable length exist which makes your case. I'm not trying to be argumentative. It's just that I can't imagine how it would be impossible to trace the exact behavior of the algorithms. It's not magic.
This is Hitch Hikers Guide to the Galaxy in real life. The answer is 42. Are you actually happy with that? (Actually the answer is a vector of 2.3 million numbers between 0 and 1, but the volume doesn't make it easier to comprehend.)
// I can't imagine how it would be impossible
I can't change the boundaries of your imagination. Interestingly, this is part of the problem - people have never encountered anything like it and so they literally cannot conceive of anything this complicated. Until I got into it I would have said the same thing.
The best I can do is the brain in the jar analogy. A complete map of every neuron firing in my head is not regarded as an acceptable explanation for why I ate eggs for breakfast. We are in a field where the limit of the possible is the neuron map.
If you dont want to believe that, the only thing left is point you to a series of math, programming, and machine learning textbooks.
I was trying to be polite, so I used "I can't imagine" and "I believe I would find." Forgive me. Your contention amounts to saying bugs cannot be fixed, nor adjustments made, because the system is too complicated. Manifestly, that is not true. Your combativeness does not strengthen your case.
// Your contention amounts to saying bugs cannot be fixed, nor adjustments made, because the system is too complicated.
No. My contention is that there are no techniques that make these systems sufficiently interpretable _today_. Maybe we will be able to do it 10 years from now, but as of this moment it simply doesn't exist.
// Manifestly, that is not true.
Electrons are both a particle and a waveform. It makes no sense, and yet it is so. Insisting otherwise doesn't change the facts.
// Your combativeness does not strengthen your case.
My apologies, I tend to avoid posting because I gravitate towards edgy analogies and snark. I'm not trying to be combative, but it falls out of me.
Using patterns of statistical association is not, to me, a "brain in a jar." Maybe we don't know every association used by a facial recognition system to rank people by beauty, but we know symmetry and feature size are key. "Too complicated" is a copout, especially for systems such as social media news and video recommendations where the results are largely predictable. Anyway, I guess this is one of those agree to disagree moments.
It has been a few years since I watched it, but Grant Sanderson (a.k.a. 3Blue1Brown) has a really nice series of four short videos on neural networks if you have the math background to follow it - mainly linear algebra and multivariable calculus. I seem to recall him doing, as he always does, an excellent job explaining why the different "layers" involved in such a network really can't be interpreted in a clear manner - even to those who designed the network.
A person's history has to be stored somewhere. A devoted Sean Hannity fan is somehow distinguishable from a Rachel Maddow devotee. In the end, it's all 1's and 0's. I believe I would find it satisfying.
That said, I'm open to being disabused should a piece of reasonable length exist which makes your case. I'm not trying to be argumentative. It's just that I can't imagine how it would be impossible to trace the exact behavior of the algorithms. It's not magic.
// I believe I would find it satisfying.
This is Hitch Hikers Guide to the Galaxy in real life. The answer is 42. Are you actually happy with that? (Actually the answer is a vector of 2.3 million numbers between 0 and 1, but the volume doesn't make it easier to comprehend.)
// I can't imagine how it would be impossible
I can't change the boundaries of your imagination. Interestingly, this is part of the problem - people have never encountered anything like it and so they literally cannot conceive of anything this complicated. Until I got into it I would have said the same thing.
The best I can do is the brain in the jar analogy. A complete map of every neuron firing in my head is not regarded as an acceptable explanation for why I ate eggs for breakfast. We are in a field where the limit of the possible is the neuron map.
If you dont want to believe that, the only thing left is point you to a series of math, programming, and machine learning textbooks.
I was trying to be polite, so I used "I can't imagine" and "I believe I would find." Forgive me. Your contention amounts to saying bugs cannot be fixed, nor adjustments made, because the system is too complicated. Manifestly, that is not true. Your combativeness does not strengthen your case.
// Your contention amounts to saying bugs cannot be fixed, nor adjustments made, because the system is too complicated.
No. My contention is that there are no techniques that make these systems sufficiently interpretable _today_. Maybe we will be able to do it 10 years from now, but as of this moment it simply doesn't exist.
// Manifestly, that is not true.
Electrons are both a particle and a waveform. It makes no sense, and yet it is so. Insisting otherwise doesn't change the facts.
// Your combativeness does not strengthen your case.
My apologies, I tend to avoid posting because I gravitate towards edgy analogies and snark. I'm not trying to be combative, but it falls out of me.
Using patterns of statistical association is not, to me, a "brain in a jar." Maybe we don't know every association used by a facial recognition system to rank people by beauty, but we know symmetry and feature size are key. "Too complicated" is a copout, especially for systems such as social media news and video recommendations where the results are largely predictable. Anyway, I guess this is one of those agree to disagree moments.
It has been a few years since I watched it, but Grant Sanderson (a.k.a. 3Blue1Brown) has a really nice series of four short videos on neural networks if you have the math background to follow it - mainly linear algebra and multivariable calculus. I seem to recall him doing, as he always does, an excellent job explaining why the different "layers" involved in such a network really can't be interpreted in a clear manner - even to those who designed the network.
https://www.youtube.com/playlist?list=PLZHQObOWTQDNU6R1_67000Dx_ZCJB-3pi
Thanks much. I'll check it out.