Research
My research is spanned by three partially-overlapping dimensions:
Neural theory to understand how the circuitry and learning rules of motor regions generate behavior.
Advanced neural and behavioral data analysis in close collaboration with experimentalists.
Development of rigorously grounded data-driven methods to identify latent structure in high-dimensional data.
I use a variety of quantitative tools from traditional methods in dynamical systems to more recent tools in machine learning.
dynamical systems
machine learning
statistical modeling
Below you can find a non-exhaustive list of my research interests.
Evolution of neural dynamics over learning
Neural circuits perform complex tasks through recurrently generated dynamics. How do these dynamic modes evolve over the course of learning, and how can we identify these changes from data recorded over learning?
Related publications:
- H Gurnani, NA Cayco-Gajic. Signatures of task learning in neural representations. Current Opinion in Neurobiology 83, 102759. (pdf)
- A Pellegrino, NA Cayco-Gajic*, A Chadwick*. Low tensor rank learning of neural dynamics. Advances in Neural Information Processing Systems 37.
Dimensionality reduction
and neural manifolds
and neural manifolds
Neural manifolds reflect task-relevant information that is often not visible in single-neuron activity. How can we develop new dimensionality reduction methods that extend the classic view that task-relevant information is embedded in low-dimensional neural subspaces while remaining (mathematically) interpretable?
Related publications:
- A Pellegrino*, H Stein*, NA Cayco-Gajic. Disentangling mixed classes of covariability in large-scale neural data. (preprint on biorxiv)
- A Pellegrino, NA Cayco-Gajic*, A Chadwick*. Low tensor rank learning of neural dynamics. Advances in Neural Information Processing Systems 37.
- F Lanore*, NA Cayco-Gajic*, H Gurnani, D Coyle, RA Silver. Cerebellar granule cell axons support high-dimensional representations. Nature Neuroscience 24, 1142-1150.
Theories of cerebellar learning
Because of its largely feedforward, relatively homogenous architecture and supervised learning signals, the cerebellar cortex has been hypothesized to perform pattern separation by projecting input patterns into a high-dimensional space, as in a multilayer perceptron. How well do these classic theories fit new experimental data? How can they be extend to understand how the cerebellum learns in coordination with other motor regions?
Related publications:
- F Lanore*, NA Cayco-Gajic*, H Gurnani, D Coyle, RA Silver. Cerebellar granule cell axons support high-dimensional representations. Nature Neuroscience 24, 1142-1150.
- NA Cayco-Gajic, RA Silver. Re-evaluating circuit mechanisms underlying pattern separation. Neuron 101 (4), 584-602.
- NA Cayco-Gajic, C Clopath, RA Silver. Sparse synaptic connectivity is required for decorrelation and pattern separation in feedforward networks. Nature communications 8 (1), 1116.
Learning during complex behaviours
Classic motor learning studies often consider simple behaviors in which variability is constrained as much as possible. But naturalistic behavior is characterized by a combination of spontaneously transitioning motor actions and continuous kinematics. How can we develop new behavioral models that can disentangle these sources of variability to understand the neural control of behavior in complex learning paradigms?
Related publications:
- A Andrianarivelo, H Stein, J Gabillet, C Batifol, A Jalil, NA Cayco-Gajic, M Graupner. Cerebellar interneuron activity is triggered by reach endpoint during learning of a complex locomotor task. (preprint on biorxiv)
- H Stein, A Andrianarivelo, J Gabillet, C Batifol, A Jalil, M Graupner, NA Cayco Gajic. The emergence of fixed points in interlimb coordination underlies the learning of stable gaits in mice. (in preparation; see an early version in Heike's Cosyne 2022 talk here)