[HN Gopher] Human-level few-shot concept induction through minim...
       ___________________________________________________________________
        
       Human-level few-shot concept induction through minimax entropy
       learning
        
       Author : geox
       Score  : 83 points
       Date   : 2024-04-19 23:54 UTC (23 hours ago)
        
 (HTM) web link (www.science.org)
 (TXT) w3m dump (www.science.org)
        
       | janpmz wrote:
       | If I understood this correctly:
       | 
       | Minimizing entropy: Learn filters that can increase the
       | likelihood of the observed data.
       | 
       | Maximizing entropy: Learning how to combine the filters.
        
         | 3abiton wrote:
         | Is this reasoning from the paper?
        
         | versteegen wrote:
         | In original FRAME, "learning filters" means selecting filters
         | from a (very small) fixed set. In this paper they add* an
         | additional "bilevel optimization", where they both learn th_j,
         | the parameters of the filters, and select the filters
         | (indicator variables z_j).
         | 
         | Either way, that learning is the entropy minimisation part.
         | 
         | Maximizing entropy is not really about learning anything, it's
         | about not overfitting. (Although you are correct in that you
         | have to find the Langrange multipliers to correctly combine the
         | filters.) Basically, the desired probability distribution will
         | be constrained only by certain statistics of (the filters you
         | learnt) on the data, and nothing else. For example, you want a
         | model that says "if panel 1 has an N-sided polygon, and panel 2
         | has an N+1-sided polygon, panel 3 has an N+2-sided polygon",
         | and says nothing about anything else.
         | 
         | * I'm a bit peeved they didn't cite me for extending FRAME by
         | learning the features like that. That was my PhD topic. But
         | let's face it, they probably don't know about my work, or
         | quickly forgot it because the papers were too ugly, and
         | admittedly they've done it in a cleaner way. I had to wrap
         | kludgy iterative feature addition around the feature
         | optimisation step to make it more efficient for image
         | modelling.
        
       | shiandow wrote:
       | It's perhaps useful to note that their notation is not
       | accidental.
       | 
       | When they state:                   P = 1/Z exp -li Hi
       | 
       | They're directly referring to the Gibbs distribution that pops up
       | in thermodynamics. The Hi are the different terms of the
       | Hamiltonian and the Z is the corresponding partition function.
       | 
       | From that point of view they're taking a couple of 'response
       | functions' treating them as if they are part of some kind of
       | energy and find the statistical distribution that such a system
       | would have at different temperatures depending on the average
       | value of each term. The next part is then to find the value for
       | each term that best predicts the observed distribution.
       | 
       | It's a bit like observing how a system behaves for different
       | volumes and then using that information to determine the
       | temerature, pressure and volume of an observed system. (perhaps
       | volume isn't the best example, but something like molarity and
       | chemical potential is hard to visualize)
        
         | versteegen wrote:
         | Yes, but...
         | 
         | > some kind of energy
         | 
         | The negenergy (negative energy) here is just the (unnormalised,
         | hence the 1/Z term) log-probability mass or density. Like the
         | logit outputs of classifier.
         | 
         | You shouldn't read anything more into the 'energy'. There is no
         | temperature parameter in these minimax models (although adding
         | one to smooth the model during sampling or learning can be
         | useful, to interpolate between pure noise and the learnt
         | model).
        
       | davedx wrote:
       | This strikes me as potentially a high impact avenue of research.
       | Inductive reasoning is core to human learning.
        
       | Hugsun wrote:
       | I can't wait to see the next innovation, that is orthogonal to
       | the scaling laws, that will cause the next leap in LLM
       | performance.
        
       | cs702 wrote:
       | Interesting. There's a lot of preprocessing involved in getting
       | these toy models to work, but the approach seems promising.
       | 
       | Is there code available, so others can examine it and replicate
       | the results? I couldn't find a link.
        
       | versteegen wrote:
       | It's surreal to see this on the front page since I worked on
       | minimax entropy learning for texture modelling (following these
       | authors), then abandoned that line of research because I wanted
       | to work on concept induction and reasoning instead. I'm amazed
       | these authors are still drawing a direct line through 27 years of
       | their papers.
       | 
       | To explain what this is about (not having read the whole paper):
       | it's using 'classical', small sample size statistics to infer
       | relations between very small sets of data such as individual
       | questions of an IQ test. Obviously that's going to give the best
       | possible answer, because that's what classical statistics does.
       | (Well, if you could solve exactly, but with this model you
       | actually can't.) Then it's just a question of having a good prior
       | over possible concepts. In this paper the concepts are hand-
       | defined parametrized features (aka 'filters' in the paper, from
       | the texture modelling lineage), e.g. "families of arithmetic
       | expressions of length from two to six" for the number problems.
       | There will be a _lot_ of details involved in that.
       | 
       | Alternatively, something like a deep neural network trained on
       | vast numbers of generated examples could come to learn what a
       | suitable prior should be by learning the patterns used to
       | generate them. Yet another option is to use a universal prior as
       | in Solomonoff induction, based on representation/coding length.
       | 
       | My conclusion: the theory is good and I'm happy to see this but
       | don't think that you can directly apply this to learning
       | 'concepts' in anything other than IQ tests. The hard part of
       | generalising this is firstly finding the prior over the concepts,
       | and secondly the difficult optimisation step: using MCMC as part
       | of the optimisation is both extremely slow and stochastic in a
       | bad way (unlike SGD) and the reason I abandoned this approach.
        
         | p1esk wrote:
         | _deep neural network trained on vast numbers of generated
         | examples could come to learn what a suitable prior should be by
         | learning the patterns used to generate them._
         | 
         | This sounds promising. Do you see any potential issues with
         | this approach?
        
           | uoaei wrote:
           | This is how NN theory describes the process of training,
           | i.e., that's the mechanism by which all neural networks come
           | to represent the problem they are being applied to solve. GP
           | is using a fancy way to say "train NNs on massive datasets"
           | which is par for the course.
        
       | eli_gottlieb wrote:
       | So they minimize the forward KL divergence from the generative
       | model to the data distribution, and then maximize the entropy of
       | the generative model? Seems like it would be simpler to just
       | minimize the reverse/exclusive KL divergence, giving you a cross-
       | entropy term to minimize and then an entropy of the generative
       | model to maximize.
       | 
       | I guess the obstacle to doing that is not being able to evaluate
       | the true data log-density, only sample from it, so the cross-
       | entropy term can't be estimated?
        
       ___________________________________________________________________
       (page generated 2024-04-20 23:02 UTC)