How to take turns: the fly's way to encode and decode rotational information

TL;DR

This paper investigates how fly H1 neurons encode rotational velocity, revealing they primarily signal velocity reversals, enabling efficient digital encoding of rotational motion with minimal information.

Contribution

It demonstrates that fly H1 neurons encode velocity reversal points, allowing simple decoding of rotational motion, a novel insight into neural encoding strategies.

Findings

Spike trains encode velocity reversals with full information.

Reversal-based encoding suffices for directional decisions.

Decoding scheme proposed for real-time stimulus reconstruction.

Abstract

Sensory systems take continuously varying stimuli as their input and encode features relevant for the organism's survival into a sequence of action potentials - spike trains. The full dynamic range of complex dynamical inputs has to be compressed into a set of discrete spike times and the question, facing any sensory system, arises: which features of the stimulus are thereby encoded and how does the animal decode them to recover its external sensory world? Here we study this issue for the two motion-sensitive H1 neurons of the fly's optical system, which are sensitive to horizontal velocity stimuli, each neuron responding to oppositely pointing preferred directions. They constitute an efficient detector for rotations of the fly's body about a vertical axis. Surprisingly the spike trains $\rho_B(t)$ generated by an empoverished stimulus $S_B(t)$, containing just the instants when the of velocity $S(t)$ reverses its direction, convey the same amount of global (Shannon) information as spike trains $\rho(t)$ generated by the complete stimulus $S(t)$. This amount of information is just enough to encode the instants of velocity reversal. Yet this suffices to give the motor system just one, yet vital order: go left or right, turning the H1 neurons into efficient analog-to-digital converters. Furthermore also probability distributions computed from $\rho(t)$ and $\rho_B(t)$ are identical. Still there are regions in the spike trains following velocity reversals, 80 msec long and containing about 3-6 msec long spike intervals, where detailed stimulus properties are encoded. We suggest a decoding scheme - how to reconstruct the stimulus from the spike train, which is fast and works in real time.