GANCAT: Generative Adversarial Networks for CATegorization

认知科学家在选择实验刺激时，往往面临着实验控制与生态效度之间的矛盾。虽然简单的刺激提供了严格的控制，但它们缺乏复杂性，在实验室和现实生活条件之间留下了解释差距。生成式对抗分类网络（GANCAT）通过开发使用新型机器学习技术的方法来帮助弥合这一差距，生成式对抗网络（GANs）用于生成复杂但完全可控的视觉刺激。GANCAT开发的方法将允许认知研究人员创造大量的自然刺激，这些刺激在实验相关的属性上有所不同。因此，GANCAT与欧盟委员会通过鼓励采用人工智能并确保人工智能系统为人民服务来实现人工智能卓越的计划保持一致。GANCAT的研究项目结合了最先进的深度学习技术和行为方法，用于研究分类、心理相似性和注意力。首先，GANCAT将复杂刺激（组织学样本）的分类与真实或gan生成的样本进行比较。其次，GANCAT结合卷积神经网络，为GAN生成的样本得出类似人类的相似性判断，并使用这些判断来识别GAN输入与在心理意义维度上变化的样本生成之间的映射。最后，GANCAT开发了一些方法，可以控制gan生成的刺激中存在的视觉显著性特征，并将这些方法用于开发加速注意力学习的自适应学习算法。GANCAT不仅有助于弥合简单和复杂视觉刺激分类之间的现有知识差距，而且还特别努力与研究界分享其工具，以说服认知科学家在他们的研究计划中欢迎现实刺激的复杂性。

When choosing experimental stimuli, cognitive scientists often face a tension between experimental control and ecological validity. While simple stimuli provide rigorous control, their lack of complexity leaves an explanatory gap between laboratory and real-life conditions. Generative Adversarial Networks for CATegorization (GANCAT) helps to bridge this gap by developing methods to use a novel machine-learning technique, Generative Adversarial Networks (GANs) in the generation of complex, yet fully controllable visual stimuli. The methods developed by GANCAT will allow cognition researchers to create large numbers of naturalistic stimuli varying across experimentally relevant properties. As such, GANCAT aligns with the European Commission's plan to achieve Excellence in AI, by encouraging AI uptake and ensuring that AI systems work for the people. GANCAT's research programme combines state-of-the-art deep-learning techniques and behavioural methods for the study of categorisation, psychological similarity, and attention. First, GANCAT compares the categorisation of complex stimuli (histology samples) as supported by real or GAN-generated samples. Second, GANCAT couples convolutional neural networks to derive humanlike judgments of similarity for GAN-generated samples and uses those judgments to identify the mapping between GAN inputs and the generation of samples that vary across psychologically meaningful dimensions. Finally, GANCAT develops methods that allow control over the visual saliency of the features present in GAN-generated stimuli and uses those methods in the development of adaptive learning algorithms that expedite attentional learning. GANCAT does not only help to bridge the existing knowledge gap between the categorisation of simple and complex visual stimuli, but it also puts special effort into sharing its tools with the research community, to persuade cognitive scientists to welcome the complexity of realistic stimuli in their research programmes.