Data locality for sparse matrices via advanced optimisations in large-scale scientific programs

通过大规模科学项目中的高级优化实现稀疏矩阵的数据局部性

基本信息

批准号：
22K17900
负责人：
Vatai Emil
金额：
$ 2.91万
依托单位：
Institute of Physical and Chemical Research
依托单位国家：
日本
项目类别：
Grant-in-Aid for Early-Career Scientists
财政年份：
2022
资助国家：
日本
起止时间：
2022-04-01 至 2026-03-31
项目状态：
未结题

项目摘要

The s-step CG topic progressed the least from the 3 topics and this direction remains mostly unexplored.: with an intern we merely made minor progress in implementing s-step CG and combining it with DMPK.We've started working on sparsity and explainability in NLP recently with another student. We have looked different approaches to sparsifycation, such as the STEN library, matrix decompositions, graph neural networks (GNNs) including graph convolutional networks (GCNs). We have gained a comprehensive understanding of the challenges of sparsity in ML training, such as sparsity crawling back quickly into our the weights (because the sum and product of matrices becomes dense easily) as well as the problem of dense gradients.Finally, the most promising direction is a collaboration with spin-chemists and the topic of radical pair simulations. We have developed a python package/library for simulating radical-pair called RadicalPy. The package includes various components which can be used to construct simulations of various experiments in the field, such as construction Hamiltonians modelling different forces in the reactions, kinetics and relaxation mechanisms, experiment schemes, data conversion and plotting tools and molecule database which enables easy and convenient loading of molecules and isotopes.

s-step CG 主题在 3 个主题中进展最少，并且这个方向大部分尚未探索。：通过一名实习生，我们在实现 s-step CG 并将其与 DMPK 相结合方面仅取得了微小进展。我们已经开始致力于稀疏性和可解释性最近和另一位学生一起学习 NLP。我们研究了不同的稀疏化方法，例如 STEN 库、矩阵分解、图神经网络 (GNN)，包括图卷积网络 (GCN)。我们对机器学习训练中稀疏性的挑战有了全面的了解，例如稀疏性快速爬回我们的权重（因为矩阵的和与乘积很容易变得稠密）以及稠密梯度的问题。有希望的方向是与自旋化学家的合作以及激进对模拟的主题。我们开发了一个名为 RadicalPy 的用于模拟根对的 python 包/库。该软件包包括可用于构建现场各种实验模拟的各种组件，例如对反应中不同力进行建模的哈密顿量、动力学和松弛机制、实验方案、数据转换和绘图工具以及分子数据库，可轻松实现方便装载分子和同位素。