My scientific work centers on extracting insights from complex neuroscience data, especially high-dimensional time-series signals recorded from the brain. During my PhD, I developed strong analytical and quantitative research skills, strengthened through coursework in computer science (machine learning track).
This is how my data look like:
I combine a deep understanding of brain physiology with advanced data science techniques, enabling me to turn raw biological signals into meaningful scientific insights.
As an initial step, I designed, trained and validated machine learning / deep learning models to investigate how cortical activity encodes movement:
Neural activity recorded from 500 cortical regions (top right) was analyzed using both a linear model and a 1D convolutional neural network (1D-CNN) (top left) to decode mouse behavioral dynamics. Although the linear model performed adequately during active movement periods, the 1D-CNN significantly outperformed it during resting (non-locomoting) states (bottom left, bottom right). These results highlight that neural encoding of behavior involves distinct nonlinearities, especially prominent during periods without overt movement
. Trained as an engineer, I believe that the best solutions are often the simplest, least expensive, and most straightforward. This principle has shaped my approach to scientific research, especially in addressing complex questions with limited resources or incomplete tools.
When I began my PhD, I was struck by a fundamental gap: standard mouse brain atlases lacked functional sub-compartmentalization of the motor cortex. Traditional structural mapping methods are often expensive, time-consuming, and labor-intensive. More importantly, they provide little information in motor areas, which are inherently functional and best studied during naturally occuring behavior.
But studying the motor cortex during movement is not straightforward. Movement is complex – not only because it involves coordinated actions across multiple body parts (e.g., hand, arm, trunk), but also because it's tightly entangled with sensation (To move is to sense). When observing cortical activity during even the simplest movement, we’re faced with a blur of overlapping signals.
Faced with this complexity, we embraced an unconventional idea: what if we could learn about motor organization not during movement, but during rest?
It turns out that during quiet, spontaneous rest, the cortex exhibits rich and structured activity patterns (see video above). These patterns are not noise – they contain information. By applying simple correlation analysis to these dynamics, I was able to quantify synchrony across the cortex and uncover distinct functional ensembles:
Seeding from well-known sensory areas (S1), I found that spontaneous activity reliably revealed corresponding motor regions with clear, body-part-specific topographies. These maps are remarkably organized – even though no overt movement is present – demonstrating that the structure of motor control is embedded in the brain’s intrinsic dynamics. These insights were only possible because of the functional lens we applied and the careful, data-driven strategies we developed to interpret them.
The result is a new, efficient approach to motor mapping – one that is low-cost, minimally invasive, and robust – and one that has uncovered organizational principles that structural methods had long obscured.
When contours from correlation maps were overlaid, it reveals a homologous spatial arrangement between sensory (S1) and secondary motor cortex (M2):
Extending this analysis by placing seed regions uniformly across posterior cortical areas revealed a continuum of functional correlation between S1/V1 and M2:
Weihao Xu, ©2025