[HN Gopher] Solving and explaining university math problems with...
       ___________________________________________________________________
        
       Solving and explaining university math problems with Deep Learning
        
       Author : dougabug
       Score  : 28 points
       Date   : 2022-08-08 16:36 UTC (6 hours ago)
        
 (HTM) web link (www.pnas.org)
 (TXT) w3m dump (www.pnas.org)
        
       | dougabug wrote:
       | New method exploits few shot learning and program synthesis to
       | automatically solve university math problems and produce
       | explanations with 10x the accuracy of previous methods.
        
       | radford-neal wrote:
       | I've only glanced at the paper, but from that glance it seems
       | that it does not "solve and explain" these problems in any thing
       | like the sense that this would mean for a human student doing the
       | problems.
       | 
       | Take the first example in Figure 4: "Find the derivative of the
       | function using the definition of a derivative. f(x) = (x**2-1) /
       | (2*x-3)". The "solution" produced is to just use a symbolic math
       | package's 'diff' function to find the derivative. I assume that
       | the actual intent of the question is for the student to use the
       | definition of a derivative: f'(x) = limit of (f(x+e)-f(x))/e as e
       | goes to zero, to find the derivative of this function, by
       | directly finding this limit.
       | 
       | The "answers" for other questions similarly miss the point. For
       | example, convergence of a series is determined by just asking a
       | symbolic math package whether it converges, not by any actual
       | reasoning, as would be expected of a student. And the question
       | asking for the Type I error probability of a statistical test is
       | "solved" using a simulation program, whereas I expect a human
       | student is expected to get the exact answer by analytical
       | calculation.
        
         | e_joules wrote:
         | See, this is a case where they went for the bombastic headline
         | and in the process hid away a really cool achievement.
         | 
         | So basically they created some sort of general purpose math
         | library, that can automatically detect the type of problem,
         | find the correct library to solve it, and input the right
         | inputs to get the right output. That is all very impressive and
         | would be a great product actually, if refined.
         | 
         | No need for the bullshit headline.
        
           | radford-neal wrote:
           | Yes, it could be useful, if seen as a natural language front-
           | end to a symbolic math package.
        
             | buescher wrote:
             | That's been done - Wolfram Alpha.
        
               | JadeNB wrote:
               | The fact that it's been done is no reason not to do it
               | again. Even if two implementations were somehow a bad
               | thing, Wolfram Alpha is the closedest of closed source,
               | and replacing it with something auditable is a win for
               | science.
        
               | buescher wrote:
               | Sure, but nobody in academia is going to get anywhere
               | describing their work as writing a replacement for Excel,
               | right?
        
           | riku_iki wrote:
           | > So basically they created some sort of general purpose math
           | library, that can automatically detect the type of problem
           | 
           | they likely trained language model on bunch of stack overflow
           | questions or something similar.
        
         | psyklic wrote:
         | The paper states it is "incapable of solving ... questions with
         | solutions that require proofs."
        
           | radford-neal wrote:
           | Then they shouldn't apply it to questions that require
           | proofs, and then claim that the system solves them when it
           | doesn't.
        
       | Bostonian wrote:
       | If a program can now solve these problems better than most people
       | can, even after studying the relevant subjects, does that mean
       | people should spend less time learning how to solve these
       | problems?
        
         | JadeNB wrote:
         | > If a program can now solve these problems better than most
         | people can, even after studying the relevant subjects, does
         | that mean people should spend less time learning how to solve
         | these problems?
         | 
         | This is a classic local-maximum situation. If you don't train
         | people to solve problems as well as a low- or middle-competence
         | program, then no-one is ever going to develop the expertise
         | that will allow them to outperform the program.
        
         | dougabug wrote:
         | I would say that people will be able to focus more on
         | understanding fundamental concepts clearly and intuitively,
         | rather than simply learning how to apply textbook math in a
         | narrow and mechanical way. I believe AI will help people absorb
         | new ideas, solidify their understanding, discover their own
         | misconceptions, and help them solve their problems without
         | getting bogged down in the details or derailed by small errors.
        
         | klyrs wrote:
         | In truth, middle-mathematics (undergrad university courses,
         | henceforth lumped into the "calculus" umbrella) is a proxy for
         | symbolic reasoning. The grungy part is looking at an arbitrary
         | problem from your area of expertise, realizing that it's
         | fundamentally a calculus problem, and then rephrasing it as a
         | calculus problem and reinterpeting the solution back to your
         | area of expertise. People have a hard time doing that if they
         | aren't familiar with the actual nuts and bolts of calculus --
         | in my experience, every student that comes out of a "black box"
         | course is entirely dependent on the specific tooling they were
         | exposed to and typically has zero curiosity about what lies
         | beneath -- those students tend to not get very far.
         | 
         | Likewise, we still teach bubble sort because it's an easy thing
         | to teach rudimentary proof skills with. Similarly, we start
         | students off with Newtonian physics.
        
         | pfortuny wrote:
         | Given GPT-3, should people spend less time learning to write?
        
       | yupis wrote:
       | This is really great. I study a lot by using the answers and the
       | steps taken to reach there.
        
       ___________________________________________________________________
       (page generated 2022-08-08 23:01 UTC)