Q&A with Cristina Savin
NYU Center for Data Science’s Cristina Savin talks brains, entropy, and computational neuroscience
Cristina Savin joined the NYU Center for Data Science (CDS) this year as an Assistant Professor of Neural Science and Data Science. She researches computational neuroscience, theoretical modelling, computer simulations, and data analysis.
1. What drew you to computational neuroscience?
I’ve always loved a good puzzle. When I was an an undergraduate studying computer science, I got interested in different aspects of artificial intelligence.
What seemed striking at the time was the huge gap between the capabilities of machines compared to humans.
It also seemed obvious that the way to make progress was to learn from the one implementation that we knew worked — namely the brain.
This brought me in contact with a computational neuroscience lab in Frankfurt, where I remained for graduate school. Over time, the more I learned about brain biology, the more my interests shifted towards neuroscience. Turns out that the mind is possibly the most complex puzzle of all.
What makes computational neuroscience intellectually challenging and really fun is the interdisciplinary and the collaborative nature of the work. My research brought me in contact with neuroscientists studying the neural mechanisms of learning in several animals, cognitive scientists studying behavior in humans subjects, and clinicians interested in impairments due to disease.
What we — the theoreticians — bring to the table is quantitative tools and new conceptual frameworks in which to think about neural computation. Success requires a constant dialogue across disciplines and this means that there is always something new to learn.
2. Can you describe a current research project that you are working on, and what role data science plays in it?
Arguably, the biggest challenge for modern neuroscience is that we are drowning in data.
Recent years have seen an explosion of experimental tools for recording and disrupting the responses of increasingly large neural populations.
The resulting data is rich, but challenging for traditional approaches (high-dimensional, non-stationary, with limited recording time).
Hence, data science is playing an increasingly important role in making sense of such data. One ongoing project in the lab tackling this problem concerns the characterization of the joint statistics of patterns of neural activity in behaving animals.
This is challenging for two reasons. First, we need to be able to estimate a complex map between the behavior of the animal and the response of individual neurons. Second, we need ways to model the statistical dependencies across neurons in a way that scales to large populations.
Our solution brings together ideas from Bayesian nonparametrics (Gaussian Processes), statistical physics (a new tractable class of maximum entropy models), and traditional frequentist statistics (hypothesis testing).
Most importantly, when using these novel data analysis tools, we were able to show that neurons in one particular area of the brain called the hippocampus work cooperatively to encode information about the position of the animal within the environment (the so-called ‘cognitive map’), and that this feature emerges as a result of learning.
3. Your research has taken you to Austria, England, France, Germany, and Romania, but it looks like it will be your first time working permanently in New York. Why do you want to pursue your research at NYU?
I’ve moved a great deal throughout my career, but I see all these moves as critical scientific stepping stones that brought my research to where it is now.
While my interest in biological learning developed during my PhD work in Frankfurt, my stay in Cambridge allowed me to hone the machine learning skills that form the core toolset used in the lab, and my stint in Vienna gave me hands-on experience with neural data analysis.
NYU, and in particular the dual appointment in neuroscience and data science, brings together the different strands of my research. It offers an unique opportunity in that it recognizes the inherently interdisciplinary nature of this research, emphasizing both machine learning and neuroscience, both theory and data analysis.
For more, see Savin’s recent co-authored paper titled “Maximum entropy models as a tool for building precise neural controls.”
Interview by Cherrie Kwok
