Tiberiu Teșileanu

Multiscale adaptation of biological systems

One of the defining features of biological systems is their impressive ability to adapt to their environments in a seemingly open-ended way. Populations of individuals can adapt to a fitness landscape through natural selection. Populations of neurons can learn to perform tasks that are far from the ones they were evolved to perform. Populations of cells in an organism can maintain homeostasis and development despite large variations in external environment.

My work aims to understand this adaptability and how it comes about. What are the fundamental principles that allow a system to adapt? What sets the limits of adaptability? How can we design systems that function optimally in uncertain environments? I look for answers to these questions using approaches rooted in physics, math, and computer science.

An example of adaptation occurs in the sensory periphery, where the structure of the brain seems adapted to the statistics of natural scenes. This “efficient coding” hypothesis has been successfully tested in vision and audition in the past. I've been working on using a similar approach to explain the distribution of olfactory receptors in the nose. A different project attempts to explain observed biases in human sensitivity to textures given the abundance of various textures in natural images. In a very different realm, modeling work that I've done in the songbird has worked out the rules that govern the adaptation between different brain regions when they are involved in two-stage learning.

Adaptation can occur across scales: for instance, individual synapses adapt through synaptic potentiation and depression, but brain function is the result of all synapses working together. This is why my work also focuses on building effective models of biological systems, that can approximate the behavior of such systems at a hierarchy of scales. One such project modeled how the CRISPR immune system in bacteria handles a viral infection. Another project involved building thermodynamically-inspired models of transcriptional regulation in eukaryotic cells. And I also work on modeling protein evolution using statistical methods and machine learning on the proteins' amino acid sequences.

In what now seems like a different life, I also worked in theoretical high-energy physics, focusing on various aspects of the AdS/CFT duality in string theory.

Neural net with nearest-neighbor interactions and STDP

each blob is a leaky integrate-and-fire neuron
redness indicates membrane voltage and spiking
neurons are noisy
there are synapses with all 8 neighbors
synapses are plastic with a timing-dependent rule
use mouse or touch to input a Gaussian-profile current into the net