One Proposal for Every Margin: Zero-Shot Amortized Sequential Importance Sampling for Binary Matrices
MarginFlow learns a zero-shot sequential importance sampling proposal, matching or beating 31 analytic designs on 1187 of 1190 margins.
MarginFlow frames sequential importance sampling proposals for binary matrices with fixed row and column sums as a generative flow network, then amortizes that proposal across margins with one set transformer. It is trained on 1904 margins and tested zero-shot on 1190 held-out synthetic and real margins from 3x3 to 870x6. On 1187 of 1190 it matches or beats the best of 31 analytic configurations chosen after the fact, with a 99.8% median effective sample fraction. On 56 hard margins it raises that median from 10.3% to 94.1%.
- Ideal SIS proposal is a GFlowNet policy with unit reward on valid matrices.
- One set transformer reads remaining margins and is reused for every margin.
- Trained on 1904 margins and evaluated zero-shot on 1190 held-out cases.
- Matches or beats the best of 31 analytic proposals on 1187 of 1190 margins.
- Median effective sample fraction is 99.8%; hard cases rise from 10.3% to 94.1%.
- Products
- MarginFlow
From the original article
1 min read at huggingface.co
In ecology, psychometrics, and the analysis of social and financial networks, binary matrices are often analyzed conditional on their observed row and column sums, which restricts the problem to a finite sample space of matrices with the same margins. Two fundamental problems are to count this space and to sample uniformly from it.
By Ruishuo Chen, Weijia Li, Xun Wang · huggingface.co
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.