Dynamic Adaptation in Fly Motion Vision

Sensory neurons process and convey information about our
surroundings, providing the physiological basis for how we interact
with the external world. In order to understand neuronal responses
we must identify the rules governing how sensory information is
encoded. It was proposed more than fifty years ago that neural
codes constitute efficient representations of the natural world
(Attneave, 1954; Barlow, 1961). In an information maximization
paradigm, an efficient coding strategy will match the encoded
neural response to the statistics of the input signals. Adaptation
of the stimulus-response function to the statistics of the stimulus
is one way to efficiently encode a stimulus when the response range
and resolution are limited compared to the entire range of stimulus
probabilities (Laughlin, 1981). Recent work has indeed shown that
adaptation to the input statistics can occur in real time
(Smirnakis et al., 1997) and that this form of adaptation can be
used to efficiently encode the stimulus and maximize information
transmission (Brenner et al., 2000). In this work I examined the
mechanisms of dynamic adaptation in fly motion vision. The H1-cell
is a large field tangential cell of the blowfly visual system that
responds to motion in a directionally selective way. It also adapts
its response properties to the second order statistics of an
apparent motion stimulus (Fairhall et al., 2001). I measured the
adaptation of the H1-cell to the variance and temporal correlations
of a Gaussian low-pass filtered velocity signal that directed a
sine wave visual grating. I found that the H1-cell adapted the
slope, or gain, and range of its input-output function to the
variance of the velocity signal over two orders of magnitude. The
H1-cell also adapted its response properties to the low-pass filter
time constant of the velocity signal over one order of magnitude. I
compared the adaptation between flies by normalizing the gain of
the stimulus-response function by the gain of the stimulus-response
function during steady-state firing properties. This “dynamic gain”
decreased as the velocity variance increased and broadened to cover
the larger range of velocities. In contrast, as the time constant
of the velocity fluctuations increased, the dynamic gain increased.
The results of these experiments were then compared with
simulations of the correlation-type or Reichardt motion detector
model. The Reichardt detector is an algorithmic model for motion
detection that explains the behavior of directionally selective
large-field tangential cells in flies including the H1-cell, as
well as directionally selective motion vision in humans (Zanker,
1996; Borst and Egelhaaf, 1989). The Reichardt detector model
showed the same adaptive properties as the H1-cell in response to
the same stimuli. Reichardt detector adaptation occurred without
changing any of the model parameters; it was an automatic function
of the dynamics of the model. This suggested that the mathematical
properties of the Reichardt detector provide a mechanism for
adaptation in the H1-cell of the blowfly. This adaptation was
further characterized in both the Reichardt detector model and the
H1-cell. The time course of this form of velocity adaptation in the
H1-cell was examined by switching between two different variances
and two different low-pass filter time constants of the velocity
signal. The H1-cell adapted to the statistics or the time course of
the new velocity signal within two seconds after the switch. The
Reichardt detector showed a similar time course for adaptation as
in the experiments. The effect of the visual pattern on adaptation
was also examined, using a square wave pattern in addition to the
sine wave used previously. The visual pattern affects the output of
an array of Reichardt motion detectors and may therefore affect
adaptation in the system. The overall shape of the adaptation
function with respect to the stimulus variance was not different
between the two stimulus patterns. In the experiments, the H1-cell
showed a consistently higher dynamic gain with a square wave
pattern. The Reichardt detector model, however, had a lower dynamic
gain when the square wave pattern was presented. After careful
investigation of the potential causes of this discrepancy I found
that the steady-state firing rate of the H1-cell saturated when a
square wave pattern was used, thereby altering the normalization
under experimental conditions that was not accounted for in the
simulations. These results suggest that contrast saturation is an
important feature of fly motion vision that has not been explained
by the Reichardt detector model. The Reichardt detector provides an
automatic mechanism and mathematical explanation for adaptation in
the fly visual system involving the nature of the incoming visual
signals and the non-linearity in the motion detector model.
Interestingly, the gradient detector model, although it is also
non-linear, does not display automatic adaptation. It remains to be
seen whether this type of adaptation is prominent in other sensory
systems and whether it leads to and efficient and accurate
representation of the natural world.