Categorization is the brain's ability to recognize the meaning of objects and events in our environment, and is an essential cognitive process underlying decision making. Categorical decisions are often flexible, and depend on the demands on the task at hand. The current project aims to understand the brain mechanisms which underlie flexible categorical decision making, as well as computational algorithms for making such decisions my artificially intelligent systems. Experiments will record from ensembles of cortical neurons during flexible categorization tasks. Computational modeling work will train recurrent neural networks to perform the same flexible categorization tasks used in the experiments, with parameters of the model inspired by the experimental data. This will result in a greater understanding of the neural mechanisms underlying categorization and decision making, as well as improvements in computational algorithms for flexible categorization by artificially intelligent systems. The broader impacts of the project include substantial training opportunities for undergraduates, Ph.D. students, and postdoctoral researchers in both experimental and computational approaches to flexible decision making. The project will also generate new experimental data and computational tools that will be shared with the broader scientific community.
This project combines multi-channel neurophysiological recordings and neural circuit modeling to investigate the neural circuit mechanisms of flexibility and generalization in visual categorization. The project leverages a collaboration by the researchers that has proven fruitful in our previous joint research on category learning. The focus of the present project is on flexible task switching between discrimination and categorization, and between categorization rules, in the behavioral, experimental, and computational work. The task paradigms will also directly test the 'exemplar model' of categorization from cognitive psychology, linking behavioral models to neural circuit processes. The project will develop a novel modeling framework, based on training recurrent neural networks to learn to perform multiple tasks. This approach offers a potentially powerful data analysis tool and conceptualization of neural circuit computation in terms of neural population trajectories in a high-dimensional state space, and this perspective is urgently needed to analyze simultaneous recording from many single neurons during performance of complex cognitive tasks, a major thread of modern Data-Intensive Neuroscience and Cognitive Science.
