A role for neural noise in animal behavior

The brain is never quiet. Activity fluctuations - or noise - impinge upon all neural circuits and may therefore have an important effect on decision-making. A former HFSP Long-Term Fellow, Pavan Ramdya, and colleagues explored this possibility by generating and studying noise-driven artificial neural networks that are able to mimic the unpredictable timing of fruit fly walking behaviors.

HFSP Long-Term Fellow Pavan Ramdya and colleagues
authored on Tue, 01 December 2015

Brain activity is not an orderly march but a complex rhythm resulting from distributed communications among connected neurons and stochastic molecular events. The resulting cacophony may have considerable impact on neural firing patterns and, consequently, animal behavior. Previous studies uncovered an important role for noise in peripheral sensing. However, the impact of noise on central neural circuits responsible for shaping moment-to-moment actions has remained largely mysterious. To gain insights into the potential role of neural noise in action timing, Ramdya and colleagues generated and studied noise-driven artificial neural networks that were able to reproduce measured patterns of animal behavior.

The researchers addressed this question using the fruit fly, Drosophila melanogaster, a powerful genetic model organism whose behavior can be studied using high-throughput assays. They measured Drosophila spontaneous and odor-evoked walking behaviors using videography and computer vision techniques. Then, using a novel algorithm and this behavioral data, they automatically generated neural networks capable of reproducing fruit fly patterns of walking and resting. They found that networks required noise inputs to accurately mimic fruit fly behavior.

Figure: A neural activity trajectory density map for a two-neuron artificial neural network capable of reproducing Drosophila walking patterns. Bottom-left indicates the lowest activity for both neurons in the network. The output neuron activity level varies across the x-axis and the hidden neuron activity level varies across the y-axis. Blue indicates regions in state space that are rarely visited and red indicates regions in state space that are frequently visited. This image was generated using data from 10,000 runs of the neural network starting from a wide range of initial conditions.

Using dynamical systems analysis they found that they each network had multistable dynamics. Noise shifted neural network activity between two stable equilibria, giving rise to short and long walking bouts similar to those seen in the fly. This strategy of using noise to transit between stable network states had the unexpected effect of increasing the potency of weak sensory stimuli on the network’s virtual behavioral output. This effect is similar to stochastic resonance observed in peripheral sensory neurons but here it emerged spontaneously to reproduce the putative activity of premotor, action-generating circuits. The researchers predict that the influence of noise on sensorimotor transformations might benefit animals as they navigate complex sensory environments.

The relatively compact and genetically accessible nervous system of Drosophila makes it possible to test these predictions by recording from and perturbing identified central, premotor neurons. Since neural noise is present across species, discoveries in the fly may also lead to a general understanding of how activity fluctuations in the brain influence decision-making and, by extension, contribute to human behavioral disorders.


Fluctuation-Driven Neural Dynamics Reproduce Drosophila Locomotor Patterns. Andrea Maesani*, Pavan Ramdya*, Steeve Cruchet, Kyle Gustafson, Richard Benton, Dario Floreano. PLoS Computational Biology DOI: 10.1371/journal.pcbi.1004577  *denotes equal contribution

Link to article

Pubmed link