Multiscale Behavior Mapping to Decode Neuronal Health and Diseases in Animal Models
Dr. Weihong Lin, Department of Biological Sciences.
Automated, high-fidelity behavioral phenotyping is essential for uncover the underlying neural circuits and functional perturbations in vivo. Yet current computer-vision tools often miss the subtle, ethologically relevant kinematic motifs that report changes in internal state. In this project, we will build a neuroscience-focused computational pipeline that couples the broad foundation model with the efficient and specialized detector to identify the animal behavior responded to sensory signal inputs. Specifically, we will distill knowledge from Vision Transformers into customized single-state Convolutional Neural Networks (CNNs) and pretrain these networks on a large, unlabeled corpus of murine behavioral videos spanning diverse common behavior experimental assays. The resulting models will be fine-tuned for high-throughput quantification of classified behavior phenotypes/signatures that serve as proxies for cognitive and affective states. This pipeline will enable experimenters to (i) stratify behavior with ethological precision, (ii) align these phenotypes with neural activities or perturbations by endogenous and environmental factors, and (iii) standardize assays across labs for reproducible neuroscience. Training and optimization demand large-scale data parallelism and extensive hyperparameter searches; thus, access to the UMBC HPC cluster is indispensable. This work will yield a biologically grounded, scalable toolset for objective behavior measurement in animal models, such as mice.