Searching for fundamental theories of Nature with unsupervised machine learning

在高能粒子物理学中，我们将数据与自然界的新理论进行对比。这些理论被提出来解决诸如1.）黑暗的宇宙是由什么组成的（2）为什么宇宙中的物质比反物质多得多; 3）为了回答这些问题，我们提出了数学模型，并与观测结果进行了比较。数据的来源相当多样化，包括大型强子对撞机的复杂测量，地下暗物质探测实验和宇宙微波背景的卫星信息。我们需要将所有这些数据整合到一个框架中，使我们能够测试假设，这通常是通过一个分析来完成的，例如贝叶斯，它提供了一个假设可以解释当前观察结果的程度。可惜，这种方法迄今为止还没有取得成果，而且正在把粒子物理学领域推向僵局。在这个项目中，我们将采取一种不同的、新颖的方法来寻找新的物理学。我们将假设我们无法发现新的物理学源于强大的理论偏见，到目前为止，这些偏见一直指导着分析。相反，我们将开发无监督的搜索技术，从数据中挖掘新现象，尽可能避免理论偏见。该项目具有很强的理论成分，因为候选人将学习新物理理论的数学/物理基础，包括暗物质，希格斯粒子和通货膨胀。候选人还将了解当前的无监督学习技术和高能物理数据的解释。该项目所采用的策略有可能为高能物理研究开辟一条新的途径。我们相信，这种与上述传统物理分析不同的方法是从大型强子对撞机和其他类似规模的实验产生的大量数据中发现新物理的最有效方法。要达到这里所概述的科学目标，需要在不同纯度水平上对大量数据进行建模。我们需要在一个新的实验中使用一个新的模型，这个新的模型将被用来分析物理学中的新物理现象。我们需要一个新的模型，这个新的模型将被用来分析物理学中的新物理现象（原始测量，伪观测，重新解释的数据），并发现由于专注于较小的信息集而未被检测到的模式。因此，我们相信在这种背景下对无监督学习的研究将具有超越学术追求的深远应用。随着世界变得越来越以数据为导向，我们也越来越依赖新的算法来理解我们所拥有的信息。举一些例子，我们可以很容易地期待将无监督学习的发展集成到面部识别软件中，并帮助发现新药，这分别为安全和医疗部门提供了推动力。

In High Energy Particle Physics we contrast data with new theories of Nature. Thosetheories are proposed to solve mysteries such as 1.) what is the Dark Universe made of, 2.)why there is so much more matter than antimatter in the Universe, and 3.) how can a lightHiggs particle exist.To answer these questions, we propose mathematical models and compare withobservations. Sources of data are quite varied and include complex measurements fromthe Large Hadron Collider, underground Dark Matter detection experiments and satelliteinformation on the Cosmic Microwave Background. We need to incorporate all this data ina framework which allows us to test hypotheses, and this is usually done via a statisticalanalysis, e.g. Bayesian, which provides a measure of how well a hypothesis can explaincurrent observations. Alas, this approach has so far been unfruitful and is driving the fieldof Particle Physics to an impasse.In this project, we will take a different and novel approach to search for new physics. Wewill assume that our inability to discover new physics stems from strong theoretical biaseswhich have so far guided analyses. We will instead develop unsupervised searchingtechniques, mining on data for new phenomena, avoiding as much theoretical prejudicesas possible. The project has a strong theoretical component, as the candidate will learn themathematical/physical basis of new physics theories including Dark Matter, the Higgsparticle and Inflation. The candidate will also learn about current unsupervised-learningtechniques and the interpretation of data in High-Energy Physics.The strategy adopted for this project holds the potential to open a new avenue of researchin High Energy Physics. We are convinced that this departure from conventional statisticalanalyses mentioned above is the most effective way to discover new physics from the hugeamount of data produced in the Large Hadron Collider and other experiments of similarscale.Reaching the scientific goals outlined here would require modelling huge amounts of dataat different levels of purity (raw measurements, pseudo-observables, re-interpreted data),and finding patterns which had not been detected due to a focus on smaller sets ofinformation. Hence, we believe that research into unsupervised learning in this context willhave far reaching applications beyond academic pursuits. As the world becomesincreasingly data-orientated, so does our reliance on novel algorithms to make sense of theinformation we have in our possession. To give some examples, we can easily expect thedevelopment of unsupervised learning integrated into facial recognition software and assistin the discovery of new drugs, which provides a boost in the security and medical sectorrespectively.