Research

The Neuroengineering Laboratory aims to reverse-engineer biological intelligence and the autonomous control of behavior. We intend to apply our discoveries toward inspiring the design of Neuroprosthetics, Robots and Artificial Intelligence (AI). Thus far, we have discovered how populations of descending brain neurons work as ensembles to control VNC motor circuits and behaviors [1]. We have also revealed how VNC interneurons ascending to the brain encode and convey ongoing behavioral state information to specific brain regions [2].

We believe that technical advances are a crucial step towards uncovering how brains works and that AI can accelerate neuroscience. To advocate for this approach, we have written review articles on the use of mathematical and engineering approaches in Neuroscience. First, we discussed how robots and neuromechanical simulations can be used to improve our understanding of motor systems [3]. Second, we explain how graph theory can be applied to study nervous system connectomes and animal behaviors [4].

Consistent with our belief in the important role of new technologies in scientific discovery, the laboratory has pioneered a number of techniques that have been crucial for reverse-engineering motor control in the fly.

First, we have developed software tools to precisely and automatically quantify behavior including DeepFly3D [5] and LiftPose3D [6], two early deep network-based approaches to infer 3D poses in behaving flies and laboratory animals. We believe that open source software can accelerate progress and democratize science. Therefore, we make all of our code publicly available on Github.

In addition to computational data analysis tools, we also pioneered the experimental and genetic (Act88F:Rpr) approaches to perform optical recordings of neural circuits in the ventral nerve cord (VNC) of behaving flies [7]. Because this approach is limited to acute recordings (up to 2-4 hours), we next worked with microrobotics colleagues at EPFL to develop microengineering tools that allow us to repeatedly record VNC circuits in the same animals multiple times over weeks [8]. These tools can allow investigators to examine how motor circuit dynamics are altered by injury, changes in internal state, and learning.

One major limitation in motor control neuroscience has been the lack of a way to synthesize findings to test their sufficiency in explaining observed animal behaviors. To address this gap, we developed the first data-driven neuromechanical model of the adult fly, NeuroMechFly [9] (code and tutorials). We also expanded upon this model to simulate full hierarchical sensorimotor control [10]. We believe that NeuroMechFly can accelerate our understanding of biological nervous systems by allowing the synthesis of findings, the exploration of experimentally inaccessible questions, and the generation of predictions for future experiments.

More recently we have begun investigating more complex coordination aspects of behavior. For example, we have used connectome-constrained simulations to study how the brain coordinates the movements of multiple body parts [11]. As well, we have begun examining how behavior can be modified through experience. In this case, we investigated the role of conspecific exposure in how animals learn to become sociable towards one another [12].