arXivDaily arXiv每日学术速递 周一至周五更新
arXiv周末暂无论文更新,休息一下吧,周末愉快~~
arXiv 2406.03495cs.LGcond-mat.dis-nnhep-thmath.NTstat.ML
收藏

模多项式任务的顿悟(Grokking)

Grokking Modular Polynomials

University of Maryland, College Park(马里兰大学帕克分校)

Darshil Doshi, Tianyu He, Aritra Das, Andrey Gromov

AI总结:

本文扩展了模加法和模乘法的解析解,组合专家解使网络能泛化到任意模多项式,并提出模多项式可学习性的分类假设及实验证据。

摘要:

神经网络能够轻易学习模算术任务的一个子集,但在其余任务上却无法泛化。这一局限性并不会因架构和训练策略的选择而改变。另一方面,文献中已知一种针对可在模加法任务上泛化的多层感知器(MLP)网络权重的解析解。在本工作中,我们(i)扩展解析解的类别,使其同时涵盖模乘法以及包含多项的模加法。此外,我们表明,在这些数据集上训练的真实网络在实现泛化(顿悟,grokking)时会学到类似的解。(ii)我们组合这些“专家”解来构建能够在任意模多项式上泛化的网络。(iii)我们提出一种假设,将模多项式按是否可通过神经网络训练学习进行分类;并提供实验证据支持我们的论断。

英文摘要:

Neural networks readily learn a subset of the modular arithmetic tasks, while failing to generalize on the rest. This limitation remains unmoved by the choice of architecture and training strategies. On the other hand, an analytical solution for the weights of Multi-layer Perceptron (MLP) networks that generalize on the modular addition task is known in the literature. In this work, we (i) extend the class of analytical solutions to include modular multiplication as well as modular addition with many terms. Additionally, we show that real networks trained on these datasets learn similar solutions upon generalization (grokking). (ii) We combine these "expert" solutions to construct networks that generalize on arbitrary modular polynomials. (iii) We hypothesize a classification of modular polynomials into learnable and non-learnable via neural networks training; and provide experimental evidence supporting our claims.

机构、总结、中文摘要均由AI分析而来,可能存在错误,请以论文正文为准。

↑