5 ms·

I don't think understanding proofs is essential to understanding AI papers. A lot of papers are "dressed up" with proofs, many of them bad. But it doesn't matt

I don't think understanding proofs is essential to understanding AI papers. A lot of papers are "dressed up" with proofs, many of them bad. But it doesn't matter if the technique can be effective.

tsumnia 7y ago
Sure, but if I'm reading a paper that starts to show equations, I can start getting lost. There's no "knowning" their garbage because it gets over-math'd. I get that's a sign of a bad paper, but my only argument for "why" is because I can't understand it. Not exactly the strongest defense. The same for stats - if you are uncomfortable with them, then seeing terms like Wilcoxon and Spearman are just confusing. Should the number be high? Low? Or simply, why do some papers have one and others have the other? If you are shaky on stats, its easier to just accept a particular formula was used even if it is not appropriate.
- codesushi42 7y ago
  My point was you can understand a paper without understanding all of the proofs shown. Following the proof isn't essential to understanding what the technique is. There are papers with bad proofs but valid techniques. Because NNs techniques are difficult if not impossible to prove mathematically. Empirical data is much more valuable.
  - tsumnia 7y ago
    Right, but my point is what if I want to implement the math? Without deep diving into my research, I disagree with your second sentence entirely. It is why I think we are also seeing the rise in Jupyter notebook still articles - because you can explicitly follow along. As a rule of thumb, I try to assume no one understands anything and so I must be very explicit in anything I teach. I find this to be a better baseline than assuming everyone already knows what I'm about to teach. Back to my point - if I want to implement something I've read in a journal article on something like queuing theory that I'd like to work into my CS education research, having a weaker foundation on math has made it harder for me to implement someone's work.
    - codesushi42 7y ago
      What you said is true, lacking the mathematical intuition will limit you. But understanding the proof is not essential to implementing the approach.
frobozz 7y ago
Academic papers in computing don't tend to be written with lay practitioners in mind. If I want to try out a technique in a business setting, I ideally want at least one of: * A library or application I can run * A code sample I can copy and modify * A pseudocode block I can turn in to real code What is less helpful is some mathematical equation in which one of the variables is the result of a technique described in another paper and a whole load of unfamiliar concepts named after people.
- codesushi42 7y ago
  Agreed. In fact, code can be more expressive than mathematical notation, and providing code would do a better job of demonstrating a theory. A lot of the proofs are just to legitimize papers. The cynic in me also thinks that by not providing code, the authors intentionally make it more difficult to reproduce results to hide tampered data. You see the same happen in science all the time with p-hacking, or with just plain fraudulent data that is purposely difficult to reproduce. Luckily there are a good number of papers that provide a repo link, but not enough yet.

I don't think understanding proofs is essential to understanding AI papers. A lot of papers are "dressed up" with proofs, many of them bad. But it doesn't matter if the technique can be effective.

tsumnia 7y ago

Sure, but if I'm reading a paper that starts to show equations, I can start getting lost. There's no "knowning" their garbage because it gets over-math'd. I get that's a sign of a bad paper, but my only argument for "why" is because I can't understand it. Not exactly the strongest defense. The same for stats - if you are uncomfortable with them, then seeing terms like Wilcoxon and Spearman are just confusing. Should the number be high? Low? Or simply, why do some papers have one and others have the other? If you are shaky on stats, its easier to just accept a particular formula was used even if it is not appropriate.

codesushi42 7y ago
My point was you can understand a paper without understanding all of the proofs shown. Following the proof isn't essential to understanding what the technique is. There are papers with bad proofs but valid techniques. Because NNs techniques are difficult if not impossible to prove mathematically. Empirical data is much more valuable.
- tsumnia 7y ago
  Right, but my point is what if I want to implement the math? Without deep diving into my research, I disagree with your second sentence entirely. It is why I think we are also seeing the rise in Jupyter notebook still articles - because you can explicitly follow along. As a rule of thumb, I try to assume no one understands anything and so I must be very explicit in anything I teach. I find this to be a better baseline than assuming everyone already knows what I'm about to teach. Back to my point - if I want to implement something I've read in a journal article on something like queuing theory that I'd like to work into my CS education research, having a weaker foundation on math has made it harder for me to implement someone's work.
  - codesushi42 7y ago
    What you said is true, lacking the mathematical intuition will limit you. But understanding the proof is not essential to implementing the approach.

frobozz 7y ago

Academic papers in computing don't tend to be written with lay practitioners in mind. If I want to try out a technique in a business setting, I ideally want at least one of: * A library or application I can run * A code sample I can copy and modify * A pseudocode block I can turn in to real code What is less helpful is some mathematical equation in which one of the variables is the result of a technique described in another paper and a whole load of unfamiliar concepts named after people.

codesushi42 7y ago
Agreed. In fact, code can be more expressive than mathematical notation, and providing code would do a better job of demonstrating a theory. A lot of the proofs are just to legitimize papers. The cynic in me also thinks that by not providing code, the authors intentionally make it more difficult to reproduce results to hide tampered data. You see the same happen in science all the time with p-hacking, or with just plain fraudulent data that is purposely difficult to reproduce. Luckily there are a good number of papers that provide a repo link, but not enough yet.