arXivDaily arXiv每日学术速递 周一至周五更新
arXiv 2005.08285math.APmath.PR
收藏

基于迭代观点的神经网络Fokker-Planck方程的严格证明

Rigorous justification of the Fokker-Planck equations of neural networks based on an iteration perspective

Jian-guo Liu, Ziheng Wang, Yuan Zhang, Zhennan Zhou

摘要:

在这项工作中,主要目标是建立神经网络的福克-普朗克方程与其微观模型之间的严格联系:扩散跳跃随机过程,它捕获了集成和发射模型中神经元集合的平均场行为。该证明基于一种新的迭代方案:通过辅助随机变量计数触发事件,随机过程的密度函数和PDE问题的解都可以用级数表示,从而大大降低了验证各子问题中密度函数与PDE解之间联系的难度.迭代方法为概率方法和偏微分方程技术的结合提供了一个通用的框架,利用它我们证明了扩散跳跃随机过程的密度函数确实是具有唯一通量位移结构的Fokker-Planck方程的经典解.

英文摘要:

In this work, the primary goal is to establish rigorous connection between the Fokker-Planck equation of neural networks with its microscopic model: the diffusion-jump stochastic process that captures the mean field behavior of collections of neurons in the integrate-and-fire model. The proof is based on a novel iteration scheme: with an auxiliary random variable counting the firing events, both the density function of the stochastic process and the solution of the PDE problem admit series representations, and thus the difficulty in verifying the link between the density function and the PDE solution in each sub problem has been greatly mitigated. The iteration approach provides a generic frame in integrating the probability approach with PDE techniques, with which we prove that the density function of the diffusion-jump stochastic process is indeed the classical solution of the Fokker-Planck equation with a unique flux-shift structure.

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。

↑