Neural dynamics and substrates of graphical knowledge

点击翻译按钮获取中文摘要

Project Summary/Abstract The ability to use explicitly structured internal models of the world is both central to biological intelligence and impaired in disorders of cognition (e.g. dementia and schizophrenia). Recent work indicates that cognitive abilities such as declarative memory and navigation rely on internal models having the structure of graphs, i.e. composed of rules and variables. Such graphical knowledge structures, or ‘schemas,’ are now thought to be what enable humans and animals alike to make extraordinary and systematic inferences, to generalize, to learn in single trials, and to imagine: thus schemas are understood to be basis of advanced cognition. Still, despite this vital importance, little is known about how neurons in the brain implement schemas. Solving this overarching problem is my scientific goal, and the aim of the present proposal. In recent work, I have discovered a collection of recurrent neural networks (RNN) – neural systems that are plausibly implemented in the brain – that can perform three cognitive tasks requiring schemas: transitive inference (TI), associative inference (AI), and identity rule inference (IRI). Importantly, these tasks are potentially more tractable alternatives to traditional schematic tasks in neuroscience: indeed, initial analyses indicate that RNNs use a solvable set of dynamical mechanisms to implement schemas, and are thus mechanistic hypotheses that have not previously existed in neuroscience. Given these findings, I hypothesize that these candidate mechanisms are used in hippocampus (HPC) and prefrontal cortex (PFC), two brain regions required for schemas. To test this guiding hypothesis, I will solve and characterize the mechanisms accomplishing AI and TI (Aim 1) and IRI (Aim 2) in RNNs (K99), then probe mechanisms experimentally via high-density recordings in the HPC and PFC of rats performing these tasks in an innovative olfactory-based paradigm (Aim 3) (R00). Achieving these Aims has the potential to establish how schemas are implemented in the brain, and therefore can clarify the neural basis of advanced cognition; this work can also clarify computational and behavioral roles of HPC and PFC, two brain structures essential to cognition. The K99 phase of this work will be done in the Zuckerman Institute for Brain and Behavior at Columbia University under the mentorship of Larry Abbott and John Cunningham, two leading authorities who will advise on neural modelling, dynamical systems, and advanced machine learning; in addition, collaborators (Drs. Stefano Fusi, Vincent Ferrera, Daphna Shohamy, Rui Costa, and Richard Axel) will contribute scientific and technical expertise on cognitive tasks (design and neurobiological interpretation) and neural recordings. The training environment also includes the wider innovative and collaborative neuroscience community at Columbia, the Columbia Center for Theoretical Neuroscience, and their associated scientific and career development opportunities. Training in the K99 phase will be crucial for both the proposed research and for establishing future scientific and professional independence as an investigator leading a research group.