Game-theoretical approach to multimodal Artificial Intelligence

This project will explore ways of incorporating game theory as the basis for building interacting deep neural network (DNN) architectures. Generative Adversarial Networks (GANs) have initiated a new area of research that uses games within more complex learning systems, allowing structured interactions among machine learning models during training. However, the focus of the current body of research in utilising game theory in DNNs around GANs has created a bias toward non-cooperative games due to their successful applications. While non-cooperative games can improve performance of individual DNNs as is evident in GANs, cooperative games can open up even more possibilities for enhancing the cognitive power of learning systems, especially when defining a competition is not possible (e.g. DNNs trained from different modalities (text, audio, video, etc.) targeting the same prediction in medical diagnosis – although each can be enhanced through a GAN, the cooperation among them is faster and more effective). Additionally, cooperative games can be easily scaled up to a large number of DNNs (hundreds or thousands), as these games are known to have the ability to involve more players. This advantage makes cooperative games a viable solution for optimising the utilisation of pre-trained models. This project will formalise the process of scaling up the cognitive power of machine learning systems, where artificial agents encapsulating DNNs interact in cooperative and non-cooperative games.