Navigation circuits in brains and intelligent machines - functional logic of key computational and memory elements.

Chair: Aurel A. Lazar. Fri. 1-3p EST

1:00-1:10p

Gwyneth Card (HHMI Janelia Research Campus)

A Functionally Ordered Visual Feature Map in the Drosophila Brain

1:20-1:30p

Large-Scale Integrated Compute-in-Memory Arrays for Reconfigurable Neuromorphic Learning and Inference

Gert Cauwenberghs (UC San Diego)

The mammalian brain offers an existence proof of remarkable general intelligence acting in complex environments, realized by hierarchical assemblies of massively parallel, yet imprecise and slow compute elements that operate near fundamental limits of noise and energy efficiency. Neuromorphic instantiations approaching such natural intelligence in very large-scale integrated circuits on custom-designed silicon microchips have evolved from highly specialized, task-specific compute-in-memory neural and synaptic crossbar array architectures that operate near the efficiency of synaptic transmission in the mammalian brain, to large tiles of such neurosynaptic cores assembled into hierarchically interconnected networks for general-purpose learning and inference. By combining extreme efficiency of local interconnects (grey matter) with great flexibility and sparsity in global interconnects (white matter), these assemblies are capable of realizing a wide class of deeply layered and recurrent neural architectures with embedded local plasticity for on-line learning, at a fraction of the computational and energy cost of digital implementations on conventional CPU and GPU hardware.

1:40-1:50p

Archimedean Property of Path Integration Mechanism

C. Randy Gallistel (Rutgers University)

The Archimedean property of a finite number system is that no matter how small the number, x, there exists an integer n such that nx >= y, for arbitrarily large y. The accumulator in path integration must have this property because the dessert ant’s odometer counts 13 mm steps to keep update distance traveled on trips at least as long a 1300 m. Thus, the odometer must count by 1’s to 100,000. This would seem to imply that the path integrating mechanism cannot be analog (e.g., a moving recurrent activity bump). On the other hand, this is not a problem for a fixed-point polynucleotide counter: Assuming only 1 bit/nucleotide, it takes only 20 nucleotides—3 for the significand (to explain the +/– 12% precision of the odometer) and 17 for the exponent. The hydrolization of 1 ATP supplies the energy to add a nucleotide; whereas, one spike in an unmyelinated axon is 10^8 ATPs, and it takes a minimum of 4 spikes/s to maintain the activity bump. Energetic and volumetric efficiency point strongly to the molecular level as the level at which the digital symbols that carry forward in time acquired information are realized in neural tissue.

2:00- 2:10p

Separation of Value and Direction Computations in Olfactory Navigation

Katherine Nagel (New York University)

In order to navigate towards goals in the environment, organisms must both compute the value of different sensory stimuli, and must represent spatial information that allows them to plan locomotor trajectories. Olfactory navigation in turbulence presents an ethological model for studying these two processes, as it requires the integration of odor value cues, that tell an organism if an odor source is worth finding, and wind direction cues, that signal the source location. Here we present evidence that these two quantities are computed in separate parallel pathways in the Drosophila brain, and integrated in local neurons of the fan-shaped body to guide navigation. We argue that this separation may allow the fly to generalize odor associations between different behavioral contexts, such as walking and flight.

2:20-2:30p

Reverse-Engineering Drosophila Action Selection and Movement Control

Pavan Ramdya (École Polytechnique Fédérale de Lausanne)

A shared goal of neuroscience and robotics is to understand how systems can be built to move effectively through the world. However, state-of-the-art algorithms for selecting and executing limbed behaviors in robots are still quite primitive compared with those used by animals. To inform robotic control approaches, we are investigating how the fly, Drosophila melanogaster, controls complex limb movements. I will discuss how we are combining 2-photon imaging of the ventral nerve cord in behaving Drosophila with physics-based simulations and neural network modeling to uncover how flies generate flexible behaviors.

2:40-3:00p

General Discussion