Generative Modelling and Representation Learning

在现代世界中，原始数据是丰富的。创建能够理解大量信息流的模型将对许多任务非常有帮助。不幸的是，大多数数据并没有整齐地打包成传统机器学习方法可以使用的格式。希望有能够从任何格式的数据中提取有用信息的方法。符合这一标准的一种学习范例是生成性建模。生成性模型查看来自特定来源的示例数据，并学习合成看起来像原始真实数据的新的假数据。合成假例子本身可能对许多应用程序都没有用处。然而，创建真实示例的能力与生俱来的是对数据结构和形式的深刻理解。因此，在这些生成性模型中，必须存在数据的“表示”，以总结数据的相关方面，如其结构和形式。这些表示法对人类和其他模型都很有用。生成性模型可以被诱使产生人类可以解释的表示形式，为我们提供大量数据的自动和广泛的汇总，这些数据超出了简单的度量标准，如均值和方差。与此平行的是，可以直接根据数据的表示来训练更多的模型，而不是原始数据本身。由于表示包含关于原始生成模型学习的数据的浓缩信息，因此通常情况下，基于表示训练的模型需要更少的总体数据来达到与直接基于原始数据训练的模型相同的性能水平。本项目旨在改进现有的生成建模技术，并制定从学习的生成模型中提取表示的新方法。产生式建模的新方法在研究界经常被提出，但我们对从数据中到底学到了什么以及如何提取这些信息的理解往往滞后。该项目将处理非常新颖的生成性建模方法，旨在提高我们对它们内部工作原理的理解。这将通过对最先进模型的实证调查以及描述它们行为的理论工作相结合来实现。该项目属于EPSRC人工智能和机器人研究领域。目前，牛津大学统计系内部正在计划进行合作。

Raw data is abundant in the modern world. Creating models that can make sense of this large flow of information would be very helpful for many tasks. Unfortunately, most data is not neatly packaged in a format that traditional machine learning methods can use. It is desirable to have methods that can extract useful information from data in any format. A learning paradigm that meets this criteria is generative modelling. Generative models look at example data from a certain source and learn to synthesize new fake data that looks like the original real data. Synthesizing fake examples may not be useful for many applications in and of itself. However, inherent in the ability to create realistic examples is a deep understand of the structure and form of the data. Therefore, within these generative models there must exist 'representations' of the data that summarize pertinent aspects of the data such as its structure and form. These representations are useful for both humans and further models. The generative models can be coaxed into producing representations that are interpretable to humans providing us with automatic and extensive summaries of large amounts of data that go beyond simple metrics such as the mean and variance. Parallel to this, further models can be trained directly on the representations of the data instead of on the raw data itself. Since the representations contain condensed information about the data learnt by the original generative model, it is usually the case that models trained on representations require much less overall data to achieve the same level of performance as a model trained on the raw data directly.This project aims to improve upon existing generative modelling techniques as well as formulate new ways to extract representations from learnt generative models. New methods for generative modelling are regularly proposed in the research community but our understanding of exactly what is being learnt from the data and how to extract this information often lags behind. The project will deal with very novel methods for generative modelling and aims to bring our understanding of their inner workings up to speed. This will be achieved through a combination of empirical investigation into state of the art models as well as theoretical work to characterise their behaviours.This project falls within the EPSRC Artifical Intelligence and Robotics research area.Collaboration is currently planned within the Oxford University Department of Statistics internally.