More real than reality: using deep learning networks to resolve how people make sense of other people's behaviour

社交互动依赖于“看到”他人行为中含义的能力：我们的朋友在打开礼物时的兴奋，在赶走蜘蛛时的厌恶，或者仅仅是他们“想要”他们目前伸手想要的东西。虽然这种对意义的归因支撑着所有的社交互动--从关于我们如何与他人建立情感关系的长期决定，到我们如何对他们的行为做出反应的时时刻刻的决定--但人们对潜在的机制知之甚少。这个项目利用人工智能/深度学习的最新进展来测试新的理论框架，这些理论框架承诺揭示这些归因背后的机制。因此，社交推论不是从对他人行为的简单解读中产生的，而是反映出积极尝试向其投射意义，测试我们对他人的看法是否反映了他们的实际行为。我们的方法基于这样的观察，即惊人相似的投影机制支撑了人工智能/深度学习方面的最新进展，用于人工2D/3D图像生成(例如，扩散模型、域自适应)，这些图像实现了良好的性能-足以愚弄人类观察者将人工图像误认为真实图像(例如，“深度假”视频)。这种偶然的机制排列使得第一次有可能在深度学习架构中实现理论上提出的机制，以测试：(1)这些机制是否确实反映了类似人类的表现和偏见，(2)这些模型是否可以像人类观察者一样，将关于所见行为的信息与外部背景信息结合起来，以及(3)它们的输出是否足够好，足以欺骗人类观察者将人造动作视频与真实的行为视频误认为是真的。这个项目将为人们显著的社会感知技能提供新的见解，并允许通过机器学习模型来测试潜在的机制。此外，它还将为人工智能/深度学习社区和英国行业提供新的方式，使当前的架构“人性化”，并确保模型与人类的期望保持一致。

Social interactions rely on the ability to "see" the meaning in others' behaviour: our friend's excitement when opening a present, their disgust when brushing away a spider, or simply that they "want" the thing they're currently reaching for. And while such attributions of meaning underpin all social interactions - from long term decisions about how we relate to others emotionally to moment-to-moment decisions about how we respond to their behaviour - little is known about the underlying mechanisms.This project leverages recent advances in AI/deep learning to test novel theoretical frameworks that promise to uncover the mechanisms behind these attributions. Accordingly, social inferences to do not emerge from a simple reading of others' behaviour, but reflect active attempts to project meaning onto it, to test whether what we think about others reflects what they actually do. Our approach rests on the observation that strikingly similar projection mechanisms underpin recent advances in AI/deep learning for artificial 2D/3D image generation (e.g., diffusion models, domain adaption) that achieve performance good-enough to fool human observers in mistaking artificial for real images (e.g., "deep fake" videos). This serendipitous alignment of mechanisms makes it possible for the firs time to implement the theoretically proposed mechanisms in deep-learning architectures, to test: (1) whether these mechanisms indeed reflect human-like performace and biases, (2) whether these models can, like human observers, combine information about the seen behaviour with outside contextual information, and, (3) whether their output is good enough to fool human observers into mistaking artifical action videos with real ones.This project will provide new insights into people's remarkably social perception skills, and allow tests of the underlying mechanisms through machine-learning models. Moreover, it will provide the AI/deep-learning community and UK industry with new ways to "humanize" current architectures and ensure that models align with human expectations.