Aurel A. Lazar
Speaker of Workshop 4
Will talk about: Neurokernal: Emulating the drosophila brain on multiple GPUs
Aurel A. Lazar is a Professor of Electrical Engineering at Columbia University. His expertise is on computing with neural circuits (in silico), and on reverse engineering the fruit fly (Drosophila melanogaster) brain (in vivo).
Dr. Lazar’s work on computing with neural circuits is centered on Neural Computing Engines and on Massively Parallel Neural Computation. He pioneered formal theoretical methods of spectrotemporal and spatiotemporal neural encoding/decoding, functional identification of spiking neural circuits and architectures and, dendritic stimulus processing and spike processing. His research group implemented massively parallel neural computation algorithms in the analog domain (graded potentials) and in the spike domain on clusters of GPUs. Code developed by these projects is available in the public domain.
Dr. Lazar’s in vivo work on Reverse Engineering the Fruit Fly Brain primarily addresses sensory processing in the early olfactory system of the Drosophila. In his sensory processing research he led a team of two graduate students who developed the first airborne odor delivery system that is both reproducible and precise; this enabled a far more accurate representation of time-varying odor stimuli in the olfactory sensory neurons and projection neurons than achievable with earlier techniques. He also initiated the Neurokernel Project, an open source platform for emulation and validation of fruit fly brain models on multiple GPUs.
The brain of the fruit fly Drosophila melanogaster is an extremely attractive model system for reverse engineering the emergent properties of neural circuits because it implements complex sensory-driven behaviors with a nervous system comprising a number of components that is five orders of magnitude smaller than those of mammals. A powerful toolkit of well-developed genetic techniques and advanced electrophysiological recording tools enables the fly's behavior to be experimentally linked to the function of its neural circuitry.
To enable neuroscientists to use these strengths of fly brain research to surmount the structural complexity of its brain and create an accurate model of the entire fly brain, we have developed an open source platform called Neurokernel designed to enable collaborative development of comprehensive fly brain models and their execution and testing on multiple Graphics Processing Units (GPUs). Neurokernel's model support architecture is motivated by the organization of the fly brain into fewer than 50 functional modules called local processing units (LPUs) that are each characterized by a unique population of local neurons. By defining communication interfaces that specify how spikes and neuron membrane states are transmitted between LPUs, Neurokernel enables researchers to collaboratively develop and refine whole-brain emulations by integration of independently developed processing units. Neurokernel will also empower researchers to leverage additional GPU resources and future improvements in GPU technology to accelerate model execution to the same time scale as a live fly brain; this will enable in vivo validation of Neurokernel-based models against real-time recordings of live fly brain activity.
We will demonstrate Neurokernel's module interfacing feature by using it to integrate independently developed models of olfactory and vision LPUs based upon experimentally obtained connectivity information.