4 ms·

To me, modern AI is just "black boxes all the way down". Even specialists don't really know what's happening. It's not encouraging or interesting. Personally,

by jokoon 4y ago

To me, modern AI is just "black boxes all the way down". Even specialists don't really know what's happening. It's not encouraging or interesting. Personally, I'm more interested in analyzing those black boxes than tinkering ones that "seems to work", would it be with graph theory, analysis, etc. To me, if something works but we're unable to really understand why it does, it's more the realm of "testing broken clocks that work twice a day". Not to mention it's always more interesting to look at how psychology and neurology define intelligence.

chongli 4y ago
I agree. I’m currently taking a neural networks course and it’s making my discomfort worse, not alleviating it. At the beginning of the course we talked about biological models of neurons and that was pretty cool, if a bit simplistic. Now we’re deep into automatic differentiation and gradient descent and a bunch of hidden layers. Ultimately it’s all just using calculus to approximate some unknown function given a sample of data. The connection to biology, to real living brains, seems like a distant memory. There is no path to understanding, from what I can see. It’s pure instrumentalism and parlour tricks.
- LeanderK 4y ago
  neural networks are vaguely inspired by the human brain but have as much in common with the humans brain as planes have with birds
- krona 4y ago
  Complex systems theory is probably a better area to at least get to grips with if you want to gain at least an idea of what the shape of our lack of our understanding is like, vis-a-vis cognition etc. Automatic differentiation is too deep in the weeds whereas the 'behaviour' of neural networks is more emergent.
- ghosty141 4y ago
  > There is no path to understanding, from what I can see. It’s pure instrumentalism and parlour tricks. Exactly like nature. It feels like emulating whats happening in nature for a very specific usecase under some constraints.
PoignardAzur 4y ago
> It's not encouraging or interesting. Well, maybe it's not encouraging, but saying it's not interesting seems like willful denial. There's a saying that in swordfighting, the goal isn't to strike or to parry or to feint, the goal is to stick the pointy end in the other guy, everything else is just a means to that end. Well, we might not understand the modern AIs' fencing technique, we have no idea why they pick certain feints, and sometimes they collapse on the ground for no reason. But on average, they are really fucking good at sticking the pointy bit in the other guy, and that's the thing that matters. We can dislike them, we can wish they were better, we can try to improve them, but one thing we can't do is ignore them. Because sooner or later, they'll be everywhere, poorly analyzed or not.
LeanderK 4y ago
> Personally, I'm more interested in analyzing those black boxes than tinkering ones that "seems to work", would it be with graph theory, analysis, etc. It's not that nobody is doing this, it's that we are making very slow advances in our understanding. There are many, many, many different angles you could take and there's so far not a clear answer which one is right. We are uncovering many interesting properties, but some just open more doors than they close. For example, the whole adversarial research (crafting inputs that are misclassified despite being similar to a correct classified example) did not lead to a quick fix but to a whole research area where we are still even trying to understand what exactly we are dealing with. It's not so much that everything is black magic and nobody cares about why the problem is really so hard, but it's instead just hard to to make progress. > To me, if something works but we're unable to really understand why it does, it's more the realm of "testing broken clocks that work twice a day". That's not fair. We have extensive empirical tests, it's really been enough time and enough eyes to see that it's not some kind of coincidence or luck but instead we really have uncovered a way to make progress we haven't dreamed about before. It's not a "broken clock works twice a day" situation, we are pretty sure of that. I understand the frustration, but I can assure you that research is totally aware of NNs pitfalls and our limitations. I can imagine that this situation is weird for someone not into ML. That to the ordinary person, the world seems pretty solved. We understand the basics of physics, chemistry etc. Now here's something where we really lack the basics, even some "newtonian physics", we use something without having a guiding theory. That's so 19th century! As someone interested in theory and empirical basics to understand the properties, I can tell you it's just really hard. Our usual mathematical tools and theories do not exactly fit, from optimization to statistical learning theory. What we do just doesn't make that much sense, but the improvements suggested by these existing theories lead to worse performance. We try to come up with new ideas and new perspective on the different problems, but so far we've not found a promising candidate. It's like early science but it's still scientific.

by jokoon 4y ago

chongli 4y ago

I agree. I’m currently taking a neural networks course and it’s making my discomfort worse, not alleviating it. At the beginning of the course we talked about biological models of neurons and that was pretty cool, if a bit simplistic. Now we’re deep into automatic differentiation and gradient descent and a bunch of hidden layers. Ultimately it’s all just using calculus to approximate some unknown function given a sample of data. The connection to biology, to real living brains, seems like a distant memory. There is no path to understanding, from what I can see. It’s pure instrumentalism and parlour tricks.

LeanderK 4y ago
neural networks are vaguely inspired by the human brain but have as much in common with the humans brain as planes have with birds
krona 4y ago
Complex systems theory is probably a better area to at least get to grips with if you want to gain at least an idea of what the shape of our lack of our understanding is like, vis-a-vis cognition etc. Automatic differentiation is too deep in the weeds whereas the 'behaviour' of neural networks is more emergent.
ghosty141 4y ago
> There is no path to understanding, from what I can see. It’s pure instrumentalism and parlour tricks. Exactly like nature. It feels like emulating whats happening in nature for a very specific usecase under some constraints.

PoignardAzur 4y ago

> It's not encouraging or interesting. Well, maybe it's not encouraging, but saying it's not interesting seems like willful denial. There's a saying that in swordfighting, the goal isn't to strike or to parry or to feint, the goal is to stick the pointy end in the other guy, everything else is just a means to that end. Well, we might not understand the modern AIs' fencing technique, we have no idea why they pick certain feints, and sometimes they collapse on the ground for no reason. But on average, they are really fucking good at sticking the pointy bit in the other guy, and that's the thing that matters. We can dislike them, we can wish they were better, we can try to improve them, but one thing we can't do is ignore them. Because sooner or later, they'll be everywhere, poorly analyzed or not.

LeanderK 4y ago

> Personally, I'm more interested in analyzing those black boxes than tinkering ones that "seems to work", would it be with graph theory, analysis, etc. It's not that nobody is doing this, it's that we are making very slow advances in our understanding. There are many, many, many different angles you could take and there's so far not a clear answer which one is right. We are uncovering many interesting properties, but some just open more doors than they close. For example, the whole adversarial research (crafting inputs that are misclassified despite being similar to a correct classified example) did not lead to a quick fix but to a whole research area where we are still even trying to understand what exactly we are dealing with. It's not so much that everything is black magic and nobody cares about why the problem is really so hard, but it's instead just hard to to make progress. > To me, if something works but we're unable to really understand why it does, it's more the realm of "testing broken clocks that work twice a day". That's not fair. We have extensive empirical tests, it's really been enough time and enough eyes to see that it's not some kind of coincidence or luck but instead we really have uncovered a way to make progress we haven't dreamed about before. It's not a "broken clock works twice a day" situation, we are pretty sure of that. I understand the frustration, but I can assure you that research is totally aware of NNs pitfalls and our limitations. I can imagine that this situation is weird for someone not into ML. That to the ordinary person, the world seems pretty solved. We understand the basics of physics, chemistry etc. Now here's something where we really lack the basics, even some "newtonian physics", we use something without having a guiding theory. That's so 19th century! As someone interested in theory and empirical basics to understand the properties, I can tell you it's just really hard. Our usual mathematical tools and theories do not exactly fit, from optimization to statistical learning theory. What we do just doesn't make that much sense, but the improvements suggested by these existing theories lead to worse performance. We try to come up with new ideas and new perspective on the different problems, but so far we've not found a promising candidate. It's like early science but it's still scientific.