The organization of neural representations for flexible behavior in the human brain

点击翻译按钮获取中文摘要

PROJECT SUMMARY Cognitive control allows us to flexibly guide our actions based on our goals. Central to most prominent theories of cognitive control is the control representation. For control to be successful, this representation is maintained in working memory by the prefrontal cortex (PFC) where it allows the same input to map to different responses depending on the context. Convergent evidence has found that the PFC encodes multiple task-relevant features of a task. However, little is known about the computational features of these control representations based on how they organize this information. This is a fundamental gap in our understanding. Here we focus on one such property, termed representational dimensionality. In technical terms, representational dimensionality refers to the number of axes needed to explain the variance in activity of a neural population across its inputs. Theoretical neuroscience has demonstrated that the dimensionality of a neural population determines a fundamental computational trade-off. A low dimensional representation will discard irrelevant information and form abstractions over its inputs. It is therefore suitable for generalization to new situations. A high dimensional representation encodes multiple mixtures of inputs into highly separable firing patterns without overlap. Understanding how generalizability and separability relate to cognitive control function promises gains on some of the most fundamental problems in control, including context-guided behavior, interference resolution, multitasking, and controlled-to-automatic behavior. The goal of this research program is to link the computational properties of high dimensional control representations to cognitive control function. Our overall hypothesis is that PFC forms high dimensional representations of task features which are needed in behavioral circumstances benefitting from separability. This hypothesis is motivated by theoretical neuroscience and foundational studies that have tested the dimensionality of PFC representations in animal models. However, no study in humans has studied high dimensional codes in PFC and no evidence in any species links dimensionality to cognitive control function. Through an NINDS R21 (NS108380), we have developed and refined two novel, complementary methods for estimating representational dimensionality from fMRI and EEG data. Using these approaches, we have found preliminary evidence that the dorsolateral PFC (DLPFC) forms a high dimensional code relative to other brain areas. We also find evidence from EEG that separability of high dimensional codes improves efficient, flexible behavior and may aid stable readout. Thus, we build on these initial observations to establish the nature, functional significance, and temporal dynamics of high dimensional control representations in the human brain.