4 ms·

Just like the conservation of energy has little effect on day to day practice of physics, besides blocking off an entire unbounded fruitless field of research i

by yters 7y ago

Just like the conservation of energy has little effect on day to day practice of physics, besides blocking off an entire unbounded fruitless field of research into free energy. In other words, impossibility results make it possible to have a practical discipline where people are focused on feasible approaches instead of spending all their time and money researching pie in the sky technologies that can never work. It's like when alchemy turned into chemistry and astrology turned into astronomy, because the fields became restricted to what is realistic. Arguably the same still has to occur in other disciplines cough AI cough.

cjfd 7y ago
Actually, conservation of energy has enormous effect on the day-to-day practice of physics.....
currymj 7y ago
AI has a ton of impossibility results. the "no free lunch theorems" are probably the best known. people now actively think in terms of designing assumptions, priors, or inductive biases to make their systems work better on the data they actually care about, because there's typically no solution that is always superior without tradeoffs. https://en.wikipedia.org/wiki/No_free_lunch_in_search_and_optimization https://en.wikipedia.org/wiki/No_free_lunch_in_search_and_op...
ProfHewitt 7y ago
With respect to Intelligent Systems, there are two important theoretical results. According to [Shamir, et. al 2019] “given any two classes C1 and C2, along with any point x∈C1, and our goal is to find some nearby y which is inside C2. Our ability to do so suggests that all the classes defined by neural networks are intertwined in a fractal-like way so that any point in any class is simultaneously close to all the boundaries with all the other classes.” Their results suggests that gradient classifiers (aka "Deep Learning") could remain forever fragile in ways that cannot be remedied by larger training sets. Also important are the results on Inference Robust Logic, e.g. speed bumps) in the following: https://papers.ssrn.com/sol3/papers.cfm?abstract_id=3428114 https://papers.ssrn.com/sol3/papers.cfm?abstract_id=3428114
- yters 7y ago
  This seems to be the case with all the widely used ML paradigms: NNs, decision forests, and SVMs. Favored for their high accuracy models which seem to come at the cost of massive overfitting and hence fragility. I used a human in the loop approach for comparison, and those models, while not as accurate, seem much more robust. I think humans are better at identifying the big picture patterns, whereas ML approaches will microoptimize their way to a fragile approximation of the big picture pattern. Practically, besides likelihood of breakdown, this means these overfit ML models are very susceptible to hacking. I predict once ML becomes more commercially widespread, we'll see a big increase in ML hacking.

by yters 7y ago

cjfd 7y ago

Actually, conservation of energy has enormous effect on the day-to-day practice of physics.....

currymj 7y ago

AI has a ton of impossibility results. the "no free lunch theorems" are probably the best known. people now actively think in terms of designing assumptions, priors, or inductive biases to make their systems work better on the data they actually care about, because there's typically no solution that is always superior without tradeoffs. https://en.wikipedia.org/wiki/No_free_lunch_in_search_and_optimization https://en.wikipedia.org/wiki/No_free_lunch_in_search_and_op...

ProfHewitt 7y ago

With respect to Intelligent Systems, there are two important theoretical results. According to [Shamir, et. al 2019] “given any two classes C1 and C2, along with any point x∈C1, and our goal is to find some nearby y which is inside C2. Our ability to do so suggests that all the classes defined by neural networks are intertwined in a fractal-like way so that any point in any class is simultaneously close to all the boundaries with all the other classes.” Their results suggests that gradient classifiers (aka "Deep Learning") could remain forever fragile in ways that cannot be remedied by larger training sets. Also important are the results on Inference Robust Logic, e.g. speed bumps) in the following: https://papers.ssrn.com/sol3/papers.cfm?abstract_id=3428114 https://papers.ssrn.com/sol3/papers.cfm?abstract_id=3428114

yters 7y ago
This seems to be the case with all the widely used ML paradigms: NNs, decision forests, and SVMs. Favored for their high accuracy models which seem to come at the cost of massive overfitting and hence fragility. I used a human in the loop approach for comparison, and those models, while not as accurate, seem much more robust. I think humans are better at identifying the big picture patterns, whereas ML approaches will microoptimize their way to a fragile approximation of the big picture pattern. Practically, besides likelihood of breakdown, this means these overfit ML models are very susceptible to hacking. I predict once ML becomes more commercially widespread, we'll see a big increase in ML hacking.