Evolution Icon Evolution
Intelligent Design Icon Intelligent Design

Dragonflies Make the Most of a Tiny Brain

Photo credit: Tony Hisgett from Birmingham, UK, CC BY 2.0 , via Wikimedia Commons.

Picture the size of a dragonfly’s head, including its eyes and brain. For such tiny assets, one would think its capabilities would be severely limited. But powered flight? Targeting and chasing and intercepting a fast-darting prey? Never underestimate the ability of biological design to pack these feats — and more — into a tiny space. The secret lies in the engineering. Engineers routinely work with optimizing performance within constraints.

In PNAS, Sara Nicholson (Flinders University, Australia) and Karin Nordström (Uppsala University, Sweden) took a look at how dragonflies and other targeting insects achieve “Facilitation of neural responses to targets moving against optic flow.” Optic flow is just what it sounds like: optical information that flows past you when you are moving. Gamers know all about simulated optic flow. They try to stay focused on their objective when information and noise is moving around them. Think “jump to light speed” in Star Wars movies, or dodging rocks as Han Solo tries to negotiate an asteroid belt to get to a planet without getting hit. Optic flow is a major issue for flying insects, and dragonflies come well equipped to deal with it — and use it.

Nicholson and Nordström consider how insects succeed at targeting prey. Consider the noisy environment involved. Get into a dragonfly’s head, like an X-wing gunner, and visualize all the clutter in the heat of battle: enemy craft on the left and right, coming in from a distance, and pieces of debris passing above and below you. Rapid correction of pitch, yaw, and roll is required for your seat and for the craft, but this only adds to the visual confusion. Insects face an even greater challenge, because their fields of view are highly textured, and their prey is small compared to the background. The prey, moreover, is constantly darting in unpredictable directions. “Efficient target detection is a computationally challenging task,” they say, “which becomes even more difficult when done against visual clutter.”

Dragonflies (order Odonata) and hoverflies (order Diptera) are among insect flyers equipped with special neurons for targeting with optic flow.

The ability of insects to successfully pursue targets in clutter is thus remarkable and suggests a high level of optimization, making the underlying neural mechanisms interesting to study. Indeed, insects that pursue targets, including predatory dragonflies and robberflies, as well as territorial hoverflies, have higher-order neurons in the optic lobes and the descending nerve cord that are sharply tuned to the motion of small, dark targets. Target-tuned neurons often have receptive fields in the part of the compound eye that has the best optics. Target selective descending neurons (TSDNs) project to the thoracic ganglia where wing and head movements are controlled, and electrically stimulating dragonfly TSDNs leads to wing movements. Taken together, this suggests that TSDNs subserve target pursuit. However, how TSDNs respond to targets moving against translational and rotational optic flow is unknown.

Keep Your Eye on the Ball

One strategy for staying on course is to keep one’s forward vision on the prey, to “lock on” to the target like fighter pilots do. (This is called “gaze stabilization.”) Another strategy is to watch for anything that moves against background. Those are helpful for initial targeting from a stationary position, but things quickly get complicated when taking flight.

However, as soon as the pursuer moves, its own movement creates self-generated widefield motion across the retina, often referred to as optic flow or background motion. In addition to self-generated optic flow, when a pursuer is subjected to involuntary deviations away from their intended flight path, for example by a gust of wind, this also generates optic flow. Quickly correcting such unplanned course deviations is essential for successfully navigating through the world.

When everything is moving — hunter, target, and background — what then? If the hunter rotates, everything in the field of vision rotates with the same angular velocity (rotational movement). If the hunter turns, by contrast, more distant objects move more slowly (translational movement). Surprisingly, many flying insects show “behavioral segregation between rotational and translational movements,” they say. “How this may influence target detection is currently not known.” Into the lab they went.

Flight Simulator

The authors put a hoverfly into a flight simulator, where they could control what kind of motion it perceived with programmed moving dots. Using electrodes, they watched the response of its TSDNs.

We found that orthogonal optic flow attenuated the TSDN target response but to a lesser degree than syn-directional optic flow. This suggests that the vector divergence between the target and the optic flow is important. Most strikingly, we found that counterdirectional optic flow increased the TSDN response to target motion, if the target moved horizontally. We found that projecting optic flow to only a small frontal portion of the eye was sufficient to elicit both TSDN attenuation and facilitation. As descending neurons control behavioral output, the response attenuation and facilitation could play a role in modulating optomotor, or gaze stabilizing corrective turns, as needed during target pursuit.

The first experiments showed that the neurons responded most strongly to counteracting optic flow: i.e., when the target was moving opposite the background optic flow. They narrowed it down further and found that frontal optic flow was “required and sufficient” to trigger TSDN response. “In summary, our results show that a small spatial window of optic flow in either the dorsal or ventral visual field is enough to strongly attenuate (Fig. 3B) or facilitate (Fig. 3C) the TSDN response to target motion.” That appears to be a clever strategy for making the most of a limited set of neurons.

This suggests that the level of vector divergence between the target and the optic flow influences the TSDN responses, so that maximum attenuation is generated at minimum vector divergence, whereas maximum facilitation is generated at maximum divergence.

Further experiments with pitch, yaw, and roll seemed to support this elegant, simple strategy. Since there are other neurons participating, though, the true picture is more complex. Further experiments altering the density of dots added some complications. Upstream small target motion detectors (STMDs) also inform the TSDNs, but in different amounts depending on the type of motion. Vector divergence from optic flow, therefore, was not enough to explain all the responses. Some neurons may inhibit other neurons in some motions but facilitate them in other motions. Further work will be required to disambiguate all the factors in play.

Nevertheless, our findings make behavioral sense. Prior to initiating target pursuit, male Eristalis hoverflies predict the flight course required to successfully intercept the target, based predominantly on the target’s angular velocity. To successfully execute an interception flight, the hoverfly turns in the direction that the target is moving. In doing so, the hoverfly creates self-generated optic flow counterdirectional to the target’s motion. In this case, the TSDNs would be facilitated, which could be beneficial. Importantly, the facilitation would take place across a range of dot densities, suggesting that even relatively sparse background textures would affect the TSDN response.

When the TSDNs are quiet, the insect can assume it is still on target with the prey. Only when contrary optic flow is perceived does the TSDN signal that a course correction is required. How head movements and wing movements factor into these rapid decisions remains to be discovered. The authors did not speculate about how these organs, neurons, and responses might have evolved.

Behold the Beast

A cover story of the IEEE Spectrum shows a magnified dragonfly head. That’s where all this processing goes on. The story, “Fast, Efficient Neural Networks Copy Dragonfly Brains,” tells how “An insect-inspired AI could make missile-defense systems more nimble.” The author, Frances Chance, works at Sandia Labs on dragonflies. Like Nicholson and Nordström., she does flight simulations but in simplified software models instead of making actual measurements on the insect’s neurons. Was a reference to evolution really necessary or helpful in her opening paragraph?

In each of our brains, 86 billion neurons work in parallel, processing inputs from senses and memories to produce the many feats of human cognition. The brains of other creatures are less broadly capable, but those animals often exhibit innate aptitudes for particular tasks, abilities honed by millions of years of evolution.

Her article contains some amazing facts: those ants in your pantry have 250,000 neurons, while dragonflies have close to a million. Dragonflies intercept and capture 95 percent of the prey they pursue. Their eyes are faster than ours, operating at “the equivalent of 200 frames per second.” Without access to GPS, a compass, or gyroscope (as far as we know), a dragonfly successfully intercepts hundreds of mosquitos per day.

What intrigues Chance is how these insects do so much with so little. The AI products that make news come at a huge processing cost. These small animals rival our best capabilities in some aptitudes, and they do it by balancing simplicity with sophistication. Her model results appear oversimplified. Maybe that is due to her assumption of evolution:

It is possible that biological dragonflies have evolved additional tools to help with the calculations needed for this prediction. For example, dragonflies have specialized sensors that measure body rotations during flight as well as head rotations relative to the body — if these sensors are fast enough, the dragonfly could calculate the effect of its movements on the prey’s image directly from the sensor outputs or use one method to cross-check the other. I did not consider this possibility in my simulation….

The simulated dragonfly does not quite achieve the success rate of the biological dragonfly, but it also does not have all the advantages (for example, impressive flying speed) for which dragonflies are known.

Frances Chance is mesmerized by the navigational achievements of insects, and glows with imagined possibilities for biomimicry. She knows she needs to check her simulation against real world dragonflies. ID advocates should encourage her to do so, because often the sophistication of biological engineering that implies designing intelligence is seen in the details.