Internal Structure of the Fly Elementary Motion Detector
Citation: Hubert Eichner, Maximilian Joesch, Bettina Schnell, Dierk F. Reiff, Alexander Borst (2011/06/23) Internal Structure of the Fly Elementary Motion Detector. Neuron (RSS)
DOI (original publisher): 10.1016/j.neuron.2011.03.028
Semantic Scholar (metadata): 10.1016/j.neuron.2011.03.028
Sci-Hub (fulltext): 10.1016/j.neuron.2011.03.028
Internet Archive Scholar (search for fulltext): Internal Structure of the Fly Elementary Motion Detector
Tagged: Neuroscience (RSS) Reichardt detector Drosophila vision motion (RSS)
Building on recent experiments showing that Drosophila motion detection splits the visual input into two parallel (L1 and L2) channels, this study further assumes that those two channels encode increases and decreases in brightness (ON and OFF). It then maps computations of directional selection to the implementation of an ON/OFF based Reichardt detector.
Under these assumptions, direction-selection is found that matches same-sign but not opposite-sign sequences, suggesting a two-quadrant (2Q) variation of the Reichardt detector. A model is proposed that both matches current observations and tries to account for past data that was interpreted as supporting four separate components to a (4Q) detector.
Goals and Methods
There are biological and philosophical reasons to imagine that a 2Q detector might be preferably to a 4Q detector: it takes half the energy and biological material, and is known to account for some of the same basic edge detection. It might also explain how such detectors avoid signed multiplication, a process considered difficult to carry out biologically. The question is whether it accounts for the full range of motion detection afforded by a 4Q Reichardt detector, and whether it matches observable responses of real flies.
Apparent motion stimuli are used for these experiments, consisting of a sequence of light increments and decrements (from light to neutral to dark and back). The H1 neuron in Calliphora and VS1-5 in Drosophila are both observed. While an idealized analysis of the data would suggest that a 4Q detector is needed, that assumes that the detectors have no information about overall brightness. In contrast, real-world neurons have some memory and some awareness of overall brightness.
As a result, a real-world 2Q detector is modelled as an array of such detectors together covering a single period of a moving sine wave grating. 6 parameters are used to tune this model: the time constant of the high-pass and low-pass filters, the DC fraction, the clip point for OFF rectification, and the synaptic imbalance, as well as a weighting of the output of the two half-detectors to mimic excitator and inhibitory synaptic transmission.
Cells are assigned a primary and null direction (PD and ND), and a 'PD-ND inversion' effect is noted in which there is a differential response based on whether a PD or ND stimulus is encountered first. This inversion is considered characteristic of real cells, and looked for in a succesful model.
Results and Analysis
The single-unit model developed is similar to a traditional Reichardt detector in many scenarios. In many it parallels the response of a Reichardt detector. It also responds to all four pairs of ON/OFF input combinations. There also remain significant differences between the modelled behavior and observed behavior. This is explained by the authors in part because firing rates 'can only decrease to 0 Hz'.
The large-scale 200-unit array of such detectors is tested against a number of inputs. It generally agrees with the expected output of a Reichardt detector. When shown a standing grating before motion begins, it shows oscillations, but when shown a moving sine grating with pseudorandom velocity profile it shows similar behavior to a Reichardt detector.
Overall the authors note a number of arguments for excluding the 2Q model, but have explanations for why each one can be neglected. They identify a few observations that argue for excluding the 4Q model, and highlight those as conclusive. Future research is needed into the presynaptic mechanism before signals get to the lobula plate, and the precise nature of the nonlinearity that takes place in the multiplication unit of the Reichardt detector (or similar model). Some specific ideas are suggested for future exploration.
Theoretical and Practical Relevance
This added to the literature trying to assess the complexity of the motion-detection mechanism of the fly. It adds weight to the claim that a fly distinguishes two, rather than all four, pairwise interactions between sequences of responses to bright and dark edges. It suggests a 2Q model with 6 parameters that can fit some past data; however this remains a question of interpretation and does not settle the question conclusively.