Parallel Computations in Insect and Mammalian Visual Motion Processing

Abstract

Sensory systems use receptors to extract information from the environment and neural circuits to perform subsequent computations. These computations may be described as algorithms composed of sequential mathematical operations. Comparing these operations across taxa reveals how different neural circuits have evolved to solve the same problem, even when using different mechanisms to implement the underlying math. In this review, we compare how insect and mammalian neural circuits have solved the problem of motion estimation, focusing on the fruit fly Drosophila and the mouse retina. Although the two systems implement computations with grossly different anatomy and molecular mechanisms, the underlying circuits transform light into motion signals with strikingly similar processing steps. These similarities run from photoreceptor gain control and spatiotemporal tuning to ON and OFF pathway structures, motion detection, and computed motion signals. The parallels between the two systems suggest that a limited set of algorithms for estimating motion satisfies both the needs of sighted creatures and the constraints imposed on them by metabolism, anatomy, and the structure and regularities of the visual world.

Figure 1

Overview of computations and circuits in fly and mouse motion detection pathways. (A) Outline of the parallel processing steps involved in motion detection in mouse retina and fly eye. Initial gain control and lateral inhibition transform absolute light levels into fractional changes in light level, or contrast. Positive and negative contrasts are divided into ON and OFF pathways, motion is computed in each, and the signals are recombined. (B) Circuit outline for the fly visual system. Photoreceptors (PRs) detect light, pass information to the neurons L1, L2, and L3, which feed into a variety of different identified relay cells before motion is computed in the neurons T4 and T5. Horizontal and vertical system (HS and VS) cells recombine the motion signals from T4 and T5. (C) Circuit outline for the mouse retina. Cone photoreceptors detect light, and their signals are passed to ON and OFF bipolar cells, which act as relays to the ON and OFF starburst amacrine cells, where motion signals are computed. Direction-selective ganglion cells integrate non-direction-selective synapses from bipolar cells and direction-selective synapses from starburst cells. For example, a rightward-preferring ganglion cell (in retinal coordinates) will collect inhibitory (GABAergic) synapses from leftward-preferring starburst neurites (that is, those that point leftward).

Figure 2

Gain control in photoreceptors. (A) Sensitivity adjusts to match the range of intensities in the visual environment. Intensity distributions may be dim (dark red curve, top) or bright (light red curve, top). Photoreceptors adjust their sensitivity to match the mean intensity and encode the intensity distribution efficiently. (B) Once adapted to a mean intensity, the photoreceptor response curve can be adjusted to best represent the intensities about that mean. The black curve (bottom) is the cumulative distribution function of the input intensities, and best encodes these intensities. The solid gray curve would not encode the tails of the distribution well, while the dashed gray curve would not make use of the full dynamic range of the response.

Figure 3

Spatial and temporal lateral inhibition. (A) The center of a spatial receptive field is antagonized by the surround region. (B) A biphasic temporal receptive field is excited by one polarity in the immediate past and antagonized by the opposite polarity in the distant past. For both spatial and temporal receptive fields, the combined positive and negative weightings tune the signals and make them sensitive to contrast transitions in space and time. The representations illustrate feed-forward mechanisms, but recurrent networks or adaptation mechanisms could also contribute to receptive field properties. (C) Tuning can change with adaptation state, notably in photoreceptors, which respond to pulses of light more quickly in bright backgrounds than in dim backgrounds. Traces show model insect photoreceptor (top) and vertebrate photoreceptor (bottom) responses to a brief flash in bright and dark backgrounds. Models of photoreceptor responses are from [156] and [45].

Figure 4

ON–OFF pathway motifs. (A) The input signal is high-pass filtered to emphasize changes in contrast, and one copy (OFF) is inverted relative to the other (ON). In the ON pathway, rectification enhances contrast increment signals; whereas in the OFF pathway, rectification enhances contrast decrement signals. (B,C) Neurons at the input to the ON and OFF pathways in the fly (B) and the mouse retina (C). In the fly, the inversion between the two pathways occurs at the second synapse after the photoreceptor. In mouse retina, the inversion of the pathways occurs at the first synapse after the photoreceptor.

Figure 5

Models of elementary motion detection. (A) A diagram of the Hassenstein–Reichardt correlator with a rightward preferred direction. Signals from photoreceptors are split, delayed, and then multiplied, before an anti-symmetric subtraction. An object moving rightward first activates the left photoreceptor a, followed by the right photoreceptor b. The delay on photoreceptor a ensures that the two signals coincide at the multiplication step (in green). The multiplication amplifies the signal. Motion in the opposite direction generates an oppositely signed signal derived from the purple multiplier. (B) A diagram of a Barlow–Levick type model with rightward preferred direction. An object moving rightward first activates photoreceptor a, which passes a positive signal through. When the motion is leftward, the delayed negative signal from photoreceptor b coincides with the positive signal from photoreceptor a and suppresses the output. (C) By expanding the nonlinearity in the Barlow–Levick model in a Taylor series (Box 1), one can rewrite the model as a series of operations with different polynomial orders. The constant term is omitted here, as are the different weightings of each expansion term. The first expansion term is just the difference between inputs, which itself is not direction-selective (DS). The next term squares the inputs before they are added; this term is also not direction-selective, since there is no nonlinear interaction in space and time. The next term multiplies the two inputs and is direction selective; this term is identical to one of the two opponent multipliers in the Hassenstein–Reichardt correlator model. The neglected terms indicated by the ellipsis contain polynomial terms of higher than second degree.