深度学习探索异质orbifold景观：Z6-II案例
133 浏览量
更新于2024-07-16
收藏 1.43MB PDF 举报
"异质混合环境中的深度学习"
在当前的科研领域，深度学习已经被广泛应用于各种复杂的任务中，如图像识别、自然语言处理和模式分析等。这篇论文"Deeplearning in the heterotic orbifold landscape"由Andreas Mütter、Erik Parr和Patrick K.S. Vaudrevange在《核物理B》期刊上发表，探讨了如何在异质（heterotic）orbifold景观中利用深度学习技术来探索弦理论模型。
异质orbifold是一种数学构造，常用于弦理论的研究，尤其是与粒子物理学的现象学相联系的部分。Z6-II是一个特定的orbifold模型，它具有双曲特性，可以用来描述可能的物理宇宙模型。作者们运用深层自动编码器神经网络来构建一个关于Z6-II orbifold景观的图表。自动编码器是一种无监督学习的神经网络，能学习输入数据的压缩表示，然后再重构原始数据。在这里，它们被用来从大量模型中学习并识别出具有潜在物理意义的模式。
尽管自动编码器在训练时并未针对Z6-II orbifold模型的现象学特性进行调整，但它依然能够识别出图表中的“肥沃岛屿”——即那些集中了大量现象学上有趣模型的区域。这些区域可能包含有希望解释现实世界物理现象的模型。随后，作者们采用决策树算法进一步分析这些图表，以提取出定义这些“肥沃岛屿”的关键属性。决策树是一种机器学习算法，能够根据一系列规则对数据进行分类，帮助研究人员理解模型之间的关联性和区别。
基于这些从图表和决策树分析中获得的信息，研究团队提出了一个新的搜索策略，旨在更有效地寻找现象学上有前景的弦理论模型。这种策略可能会极大地推进对弦理论及其在粒子物理学中的应用的理解，尤其是在寻找可能与标准模型兼容的理论模型方面。
这篇论文展示了深度学习如何在理论物理研究中发挥重要作用，特别是在处理高维度、复杂数据集的背景下，能够帮助科学家发现和理解隐藏在数据背后的模式和结构。通过这种方法，未来的研究可能会揭示更多关于宇宙基本构造的新洞察。这篇论文是开放获取的，意味着公众可以自由访问和使用其研究成果，促进了科学知识的传播和交流。
116 A. Mütter et al. / Nuclear Physics B 940 (2019) 113–129
lead to fertile islands in the string landscape, i.e. to patches in the parameter space of Z
6
-II
models where the number of MSSM-like models is above average.
Let us start with an ove
rview of the main points of the following discussion. We start with
the preprocessing of our data, where we transform each Z
6
-II model into a suitable, machine-
readable representation of 26 parameters X, also known as features. Then, we utilize a neural
network to project each Z
6
-II model to a point in a two-dimensional image, yielding a “chart” of
the Z
6
-II landscape. This is done such that the reconstruction error (i.e. the error when we map
each point of the two-dimensional chart back to a feature vector X) is as small as possible. In
this chart of the Z
6
-II landscape we can easily identify fertile islands where MSSM-like models
appear to cluster – even though the neural network had no information of a model being MSSM-
like or not during training. Afterwards, a decision tree is used to investigate these fertile islands,
i.e. to find conditions on the 26 features X of a Z
6
-II model, such that one can directly decide
if a given Z
6
-II model is located on a fertile island of the landscape or not. Finally, we discuss
the performance of this procedure: we analyze how many MSSM-like models can be found if we
restrict ourselves to search for MSSM-like models only on the fertile islands.
3.1. Data preprocessing
We start our machine learning workflow with the most basic, but crucial step: to define our
training and validation sets. The training set is used in the machine learning algorithms to actu-
ally tune the weights and biases in the neurons, while the validation set is used to estimate the
generalization properties of our machine learning model and can be e
xploited for hyperparameter
search, e.g. to adjust the architecture of the neural network. Both of these sets contribute to the
structure of the machine learning model.
In our case, we ha
ve a coarse sample of O(700, 000) Z
6
-II models. This dataset is used to
build our machine learning algorithm and is divided into 60% training and 40% validation data,
all in a random procedure.
In order for the autoencoder to handle the data, we need a suitable numerical representation
of the data. In our case, there e
xists a natural representation: the 26-dimensional feature vector of
integers X, see Appendix A. However, it turns out that this representation does not perform well
on the autoencoder. In f
act, a more abstract representation, a so-called one-hot encoding, leads to
a much better result. One-hot encoding is an approach for data that has no internal order like the
values “green”, “red”, “blue”. It generates a vector with n components where n equals the total
number of possible values. Hence, in the e
xample of three colors we have n = 3 and “green”,
“red” and “blue” have a one-hot encoding (1, 0, 0), (0, 1, 0) and (0, 0, 1), respectively. In our
case of Z
6
-II models, each feature X
k
of X can take 37 different values (i.e. there are in total 37
different breaking patterns for each E
8
factor). Thus, each component X
k
of the 26-dimensional
feature vector X is represented by a 37-dimensional vector. This 37-dimensional vector is zero
except for the component, which corresponds to the given value of X
k
. This component equals 1.
Therefore, we obtain for each Z
6
-II model a (26 × 37 = 962)-dimensional feature vector X
one-hot
as input to our neural network.
3.2. The autoencoder
The main effect of an autoencoder neural network is that redundancies in the feature vec-
tor X
one-hot
(such as irrelevant features) can be detected and reduced. Thus, an autoencoder
剩余16页未读，继续阅读
134 浏览量
6412 浏览量
154 浏览量
2021-09-07 上传
2021-09-27 上传
2021-08-12 上传
104 浏览量
683 浏览量
126 浏览量
weixin_38744153
- 粉丝: 348
- 资源: 2万+
我的内容管理
展开
最新资源
- -ignite-template-corrigindo-o-codigo
- 初级java笔试题-earthshape:从天文观测重建地球形状的程序
- 店长的定位
- smzdm_checkin_daily:「什么值得买」自动签到脚本
- gleam_parser:Gleam中的解析器组合器库，深受elm-parser的启发
- Event-Organiser:一个Kotlin应用程序来组织您的活动
- 初级java笔试题-termite:终极实时策略
- Giá Hextracoin-crx插件
- utility-ThreadPool-ios:自1.2版以来，Lightstreamer的iOS客户端库使用的线程池和URL调度库
- GIS-colouring-graph-vertexes:一个 GIS 项目，其任务是实现一种算法，该算法使用相似矩阵为图形顶点着色
- AFC代码：马里兰大学量子内存实验的代码库
- Метки для учебника javascript.ru-crx插件
- 斑马官方驱动XP系统.rar
- tesseract_example:CPPAN的非常基本的Tesseract-OCR示例。 Cppan支持已终止。 请改用sw（cppan v2）。 更新的示例在这里
- OrigamiProject3
- django-mongodb-sample-login:使用Rest Freamework的Django mongodb示例应用程序
