Generative Modelling for Sequenatial Human Behaviour

这项研究的目标是分析和模拟连续的人类行为，如语言、面部动画和手势。实现这一目标的第一步是确定促成和指导人类行为的因素。生成模型可以揭示数据中的隐藏结构，这些结构通常被称为数据的潜在表示。作为我研究的一部分，我想专注于寻找解开这些潜在变量的方法，以便获得对生成序列的不同方面的控制。我想探索最近的生成方法，并使它们适应于处理顺序数据。其中一个模型是生成对抗网络，它使用一个判别网络来驱动生成网络的学习。这种方法在生成静态数据方面非常成功，将其扩展到序列是一个非常活跃的研究领域。探索这种方法的另一个优点是，它允许使用多个鉴别器网络，能够同时捕获真实人类行为数据的各个方面。这项研究的另一个目标是了解导致人类行为的信号之间的关系，因为这些信号通常是相互关联的。人类语言和面部动画之间的关系就是一个很好的例子。我计划利用这种关系来制作语音驱动动画，这将大大降低计算机生成图像（CGI）的成本。此外，我认为重要的是研究模拟信号变化的方法，而不是信号本身。这样的模型可能更适合于捕捉大多数自然系统的动态，因此我想研究新的网络架构，使这些方法成为可能。最后，当我们寻求不断提高生成的人类行为数据的真实性时，找到区分生成信号和真实信号的方法也很重要。生成非常真实的内容可能会有很大的安全隐患（即身份盗窃），因此我也想探索区分真实和生成的人类行为数据的方法。这项研究符合EPSRC在计算机图形学和人机交互领域的目标，因为它将使动画角色的快速高效生成成为可能。

The goal of this research is to analyze and model sequential human behaviour such as speech,facial animation and gestures. The first step towards achieving this is to identify the factors thatcontribute and guide human behaviour. Generative models can reveal hidden structures in thedata which are often referred to as the latent representation of the data. As part of my researchI want to focus on Finding ways to disentangle these latent variables in order to gain control overdifferent aspects the generated sequences.I would like to explore recent generative methods and adapt them to handle sequential data.One such model is generative adversarial networks, which uses a discriminating network to drive thelearning of a generating network. This approach has been very successful for generating static dataand its extension to sequences is a very active area of research. Another advantage of exploringthis approach is that it allows the use of multiple discriminator networks that are capable ofsimultaneously capturing various aspects of real human behavioural data.Another goal of this research is to understand the relationship between the signals that make uphuman behaviour because these signals are often linked. A good example of this is the correlationbetween human speech and facial animation. I plan to exploit this relationship to perform speech-driven animation which will greatly reduce the cost of computer generated imagery (CGI).Additionally, I believe that it is important to research methods that model the changes in signalsrather than the signals themselves. Such models may be better suited to capture the dynamics ofmost natural systems, hence I would like to research new network architectures that make theseapproaches possible.Finally as we seek to constantly improve the realism of generated human behavioural data itis also important to find ways to distinguish generated signals from real ones. Generating contentthat is very realistic can have great security implications (i.e. identity theft) and therefore I wouldalso like to explore ways of distinguishing real and generated human behavioural data.This research is in line with the goals of EPSRC in the fields of computer graphics and human-computer interaction since it will enable the fast and efficient generation of animated characters.