This body of work presents evidence that the minimal unit of brain state is much shorter than the time scales associated with traditional oscillatory patterns—delta, theta, alpha, beta, and gamma waves. We demonstrate that shorter time scales are sufficient by showing that raw data segments down to 40 ms carry useful information to resolve brain state, and 400 ms segments do so without substantial loss of accuracy. Employing long-term, high-fidelity electrophysiological recordings from freely behaving mice, we provide evidence that brain state is not purely a global phenomenon. Rather, individual circuits act independently of the whole for brief periods on the order of tenths of seconds, correlating with certain behaviors, such as brief pauses in wakefulness, as well as nighttime twitches and movements. These findings challenge entrenched notions that waveforms are fundamental descriptors of brain state and that brain state is globally synchronized—concepts deeply ingrained in conventional analytical methods, representing an investigative bias applied over decades. Overcoming this bias by employing modern neural network models is a cornerstone of this thesis. Although neural networks are widely criticized for being hard to interpret, we employ them in computational ablative studies by systematically modifying the input to observe how the models perform under varying conditions. This approach allows us to draw logical conclusions through extensive analysis. Recognizing that neural networks excel at processing raw signal data, we applied similar techniques to quasar spectra from the Sloan Digital Sky Survey, which are structurally analogous to electrophysiological recordings. This interdisciplinary approach enabled us to develop a model that informed our core neuroscience findings. We processed raw electrophysiology and signal data with compute-intensive neural network models, which required substantial computational resources. Utilizing the UC-run National Research Platform—a highly distributed compute cluster—we processed tens of terabytes of data. This significant computational effort was essential for handling high-volume data efficiently and contributed to the development and publication of a scalable architecture for large-scale electrophysiology recordings.
 

Event Host: David Parks, Ph.D. Candidate, Biomolecular Engineering & Bioinformatics

Advisor: David Haussler