Frozen 1-RSB structure of the symmetric Ising perceptron |
| |
Authors: | Will Perkins Changji Xu |
| |
Affiliation: | 1. School of Computer Science, Georgia Institute of Technology, Atlanta, Georgia, USA;2. Center of Mathematical Sciences and Applications, Harvard University, Cambridge, Massachusetts, USA |
| |
Abstract: | We prove, under an assumption on the critical points of a real-valued function, that the symmetric Ising perceptron exhibits the ‘frozen 1-RSB’ structure conjectured by Krauth and Mézard in the physics literature; that is, typical solutions of the model lie in clusters of vanishing entropy density. Moreover, we prove this in a very strong form conjectured by Huang, Wong, and Kabashima: a typical solution of the model is isolated with high probability and the Hamming distance to all other solutions is linear in the dimension. The frozen 1-RSB scenario is part of a recent and intriguing explanation of the performance of learning algorithms by Baldassi, Ingrosso, Lucibello, Saglietti, and Zecchina. We prove this structural result by comparing the symmetric Ising perceptron model to a planted model and proving a comparison result between the two models. Our main technical tool towards this comparison is an inductive argument for the concentration of the logarithm of number of solutions in the model. |
| |
Keywords: | frozen solutions learning algorithms perceptron planted model solution space |
|
|