Fly motion vision is based on Reichardt detectors regardless of the signal-to-noise ratio
Citation: J Haag, W Denk, A Borst (2004/11/01) Fly motion vision is based on Reichardt detectors regardless of the signal-to-noise ratio. PNAS (RSS)
DOI (original publisher): 10.1073/pnas.0407368101
Semantic Scholar (metadata): 10.1073/pnas.0407368101
Sci-Hub (fulltext): 10.1073/pnas.0407368101
Internet Archive Scholar (search for fulltext): Fly motion vision is based on Reichardt detectors regardless of the signal-to-noise ratio
Tagged: Neuroscience (RSS) vision algorithms (RSS)
It has been theorized that fly vision might switch from relying on Reichardt detectors to relying on some other mechanism such as gradient detectors in high signal-to-noise regimes. This paper summarizes two experiments probing those regimes and testing different parts of the fly vision chain, and finds no evidence for any mechanism other than Reichardt detectors.
Goals and Methods
Gradient detectors are still considered a competing mechanism to Reichardt detectors in high signal-to-noise regimes. And in order to work with elementary responses, experiments often work at the opposite end of the signal spectrum, aiming for the smallest signal that elicits a measurable response. As a result few experiments have been carried out to test whether markers of gradient detection can be observed in fly motion vision under any circumstances.
This paper describes two experiments to evaluate the hypothesis that gradient detectors are not part of the fly motion pathways at all. Both study direction selectivity in flies, the canonical test for fly motion vision.
One experiment tested for the dependence of optimal stimulus 'velocity' (the actual velocity of the pattern used as a stimulus) on the wavelength of the pattern. This is expected to be directly correlated in the Reichardt detector case, and relatively uncorrelated in the gradient detector case. It was simply measured by looking at neuron spikes of the motion-sensitive neuron H1. This test was repeated over a wide range of mean luminance and of stimulus contrast (roughly 2 magnitudes in each case).
The second more complex experiment improved on a traditional experiment to show Reichardt detection: observing the local modulations in signal along a dendrite as a pattern moves past it. In the Reichardt case this is expected to move synchronous with the pattern, and to be phase-shifted along different parts of the dendrite. However this was always previously done with 1-photon imaging, which has a side-effect of false positives that limits how small overall noise could be in such an experiment (and so limits the maximum pattern contrast). This is addressed here by using 2-photon microscopy with luminescent Calcium markers, a technique that allows very high signal without false-positive noise. Neurons of both vertical and horizontal systems were observed.
These experiments benefitted from a clear understanding of fly biology, and did not measure conscious motion detection, but neuron-level detection within the lobula plate. The flies had their heads opened and trachea and air sacs removed so that the lobula plate could be imaged directly from above; those subject to the higher SNR-requirement two-photon microscopy also had their proboscis and gut removed (they produce occasional reflexive signals). Images were taken with a 64x64 pixel camera and transformed from fuorescence changes into projected DC + AC signal.
Results and Analysis
Results at all signal levels matched what would be expected of Reichardt detectors, and while there were some variations in response at different levels of signal intensity, no hallmarks of gradient detection were seen.
This held true up to pattern contrasts of ~90% and luminance of 200cd/m^2, close to broad daylight. Additionally, the total variation in [Ca2+] at the dendritic fringe increased with increasing contrast, the opposite of what might be expected with a gradient detector.
This demonstrated that Reichardt detectors seem to model the mechanism of direction selectivity in fly neurons of the lobular plate, up to fairly high contrast and illumination levels, close to the maximum that flies are thought to distinguish. While it cannot rule out the possibility of a mechanism like a gradient detector in use in the visual system, this is a strong indication that gradient detectors are not needed for effective motion detection.
Three outstanding questions are noted as remaining support the idea that something other than Reichardt detectors, perhaps gradient detection, is involved in some regimes. One is theoretical: a naive analysis suggests that gradient detectors would be the simplest, and perhaps most efficient, detection scheme. If this is not true in any regime, where does the simple model fail? And a few experimental results are also unexplained:
- Reichardt detectors have a quadratic relationship between signal amplitude and stimulus contrast. But this becomes contrast-independent slightly above 10% contrast, in Drosophila, Musca, and others.
- Humans perceive low-contrast gratings to be slower than high-contrast gratings.
- Srinivasan (1991) notes free-flying honey bees moving through a tunnel can detect actual image velocity, not just pattern wavelength.
- Visual summary (presentation form, pdf)
- M. V. Srinivasan, M. Lehrer, W. H. Kirchner, S. W. Zhang (1991) Visual Neuroscience 6.
Theoretical and Practical Relevance
This is a concise demonstration that a Reichardt-detector process, or something very close to it, continues to be the dominant way gradients and edges are detected in fly vision, even in environments with low noise.
Historically, gradient detection has been a top contender for a mechanism for vision in low-noise environments, because of their theoretical simplicity and high precision. This experiment tried two unrelated methods of finding indications of gradient detection in fly vision, removing much more noise from their method and increasing the signal dramatically beyond previous experiments; but without success.