The Theory of Cognitive Systems Group from Bochum visited the Developmental Dynamics Lab at UEA to work on a grant project funded by the Leverhulme Trust entitled: “Toward a whole-brain, cross-domain theory of flexible cognition”. The goal of this project is to understand how humans integrate lower-level and higher-level cognitive processes to achieve flexible, real-world cognition.
We are pursuing this goal by developing a new model called WOLVERINE (Word Object Learning via Visual ExploRation In Natural Environments). Core goal 1 is to show that WOLVERINE is able to simulate both lower-level cognition (e.g., attention) and higher-level cognition (e.g., word learning). We will demonstrate this by simulate findings from canonical laboratory tasks, showing that the same model can achieve good fits to data from various paradigms.
Core goal 2 is to demonstrate real-world functionality. Here, we will show that WOLVERINE can handle real-world inputs in a robotics scenario, attending to various objects via multiple saccades and learning from a teacher in real-world word-learning scenarios.
We made excellent progress on building ‘VERINE’ 1.0 – this is a model of visual exploration that integrates prior work on visual attention and visual working memory with a saccade system that visually explores objects in its spatiotopic visual field. We also discussed expansion of VERINE to add word-object learning. This will include integration with a new model of auditory word form learning that we are developing in parallel with WOLVERINE.
Another project being pursued with the grant is to develop an automatic optimization method for neural field parameters. A common task when building neural process models with DFT is to find suitable parameters for the neural dynamics that make it to behave in accordance with empirical constraints. Specifically, fields have to be in the right dynamic regimes, and the coupling between fields has to be right. Building on established state-of-the-art methods for optimizing connectionist or deep recurrent neural networks, we have developed a novel method for automatically determining suitable parameters for the time-continuous neural dynamics of DFT. We have demonstrated that this method is able to find parameters that allow fields to be in the right dynamic regimes such as to satisfy various empirical sources of constraints, including reaction times.