The brain is a complex dynamical system, with a hierarchy of spatial and temporal scales ranging from microns and milliseconds to centimeters and years. Activity at any given scale contributes to activity at the scales above it and can influence activity at smaller scales. Thus a true understanding of the brain requires the ability to understand how each level contributes to the system as a whole. Most brain research focuses on a single scale (single unit firing, activity in a circuit), which cannot account for the constraints imposed by activities at other scales. The goal of this proposal is to develop a framework for the integration of multiscalar, multimodal measurements of brain activity. One of the challenges in understanding how activity translates across scales is that features that are relevant at one scale (e.g., firing rate) do not have clear analogues at other scales. We address this issue by defining trajectories in ?state space? at each scale, where the state space is defined by parameters and time scales appropriate to each type of data. The trajectory of brain activity through state space can uncover features like attractor dynamics and limit cycles that characterize the evolution of activity. Using machine learning along with new and existing multimodal measurements of brain activity (MRI, optical, and electrophysiological), we propose to establish methods that relate trajectories across scales while handling the mismatch in temporal sampling rates inherent in multi-scale data.
Specific aims are 1. Create and test a tool for learning how trajectories at fast scales influence activity at slower scales. Different modalities have different inherent temporal resolutions in addition to different types of contrast. Current methods generally downsample the faster modality in some way, losing much information in the process. We will leverage variants on long short-term memory (LSTM) network architectures to learn the relationship between state space trajectories acquired simultaneously with population recording and optical imaging, and with optical imaging and fMRI. 2. Create and test a tool for learning how trajectories at slow scales influence activity at faster scales. Leveraging the same LSTM-based approach, we will learn how slower, larger scale activity affects activity at smaller scales, using whisker stimulation as a test case. We anticipate inclusion of the large scale activity (measured with fMRI or optical imaging) will improve prediction of the response at smaller scales (measured with optical imaging or population recording). Our work will allow us to begin to answer a wide range of questions about how the brain functions (e.g., what type of localized stimulation that will drive the brain to a desired global state? How does modulation of the global brain state affect local information processing?) and provide guidance for future experiments by identifying key features that influence activity across scales. By approaching the whole brain as a complex dynamical system, we will break free from the limitations of previous studies that focus on individual cells or circuits. We also expect our work to stimulate new theories that incorporate multiple scales of activity.
The brain is a complex, dynamical system that exhibits structured activity at many different space and time scales. Most existing studies focus on a single scale (for example, spiking of individual neurons or activity in a particular circuit). However, interactions occur across scales and may account for much of the variability observed in the brain. This proposal will develop a framework for integrating data from different scales using machine learning tools. The results will improve our understanding of how the brain functions and serve as a foundation for modeling neuromodulatory treatments for brain disorders.
