Neural and sensory systems adapt to prolonged stimulation to allow signaling across broader input ranges than otherwise possible with the limited bandwidth of single neurons and receptors. In the visual system, adaptation takes place at every stage of processing, from the photoreceptors that adapt to prevailing luminance conditions, to higher-order motion-sensitive neurons that adapt to prolonged exposure to motion. Recent experiments using dynamic, fluctuating visual stimuli indicate that adaptation operates on a time scale similar to that of the response itself. Further work from our own laboratory has highlighted the role for rapid motion adaptation in reliable encoding of natural image motion. Physiologically, motion adaptation can be broken down into four separate components. It is not clear from the previous studies which of these motion adaptation components are involved in the fast and dynamic response changes. To investigate the adapted response in more detail, we therefore analyzed the effect of motion adaptation using a test–adapt–test protocol with adapting durations ranging from 20 ms to 20 s. Our results underscore the very rapid rate of motion adaptation, suggesting that under free flight, visual motion-sensitive neurons continuously adapt to the changing scenery. This might help explain recent observations of strong invariance in the response to natural scenes with highly variable contrast and image structure.
Neural and sensory systems adapt to prolonged stimulation, thereby optimizing sensitivity to novel stimuli. In the visual pathways, adaptation takes place at all stages: from the photoreceptors themselves, through several post-receptor stages (e.g. Kim and Rieke, 2001; Laughlin, 1989; Rieke and Rudd, 2009; Smirnakis et al., 1997), to higher-order motion-sensitive visual neurons in the brain (e.g. de Ruyter van Steveninck et al., 1986; Kohn and Movshon, 2003; Maddess and Laughlin, 1985; Solomon et al., 2004). Psychophysically, adaptation to prolonged motion stimulation includes the classic ‘waterfall effect’ for human observers (e.g. Clifford and Ibbotson, 2002). Motion adaptation has been studied extensively in the vertebrate visual cortex, as well as in fly lobula plate tangential cells (LPTCs) (e.g. Kohn and Movshon, 2003; Kurtz et al., 2000; Maddess and Laughlin, 1985; Price et al., 2006; Reisenman et al., 2003). Direction-selective LPTCs are readily identified based on their physiological response properties, allowing data to be pooled from recordings in different animals, thus enabling us to quantify motion vision in more detail than possible in cortical recordings. Importantly, several studies support similar mechanisms underlying motion adaptation in flies and vertebrates (e.g. Clifford and Langley, 1996), making insects an excellent model system for quantifying the underlying neuronal processes.
Motion adaptation can be broken down into four separate components (Harris et al., 2000; Kohn and Movshon, 2003; Nordström and O'Carroll, 2009): (1) a non-directional contrast gain reduction, (2) an antagonistic (direction-selective) after-potential, (3) a non-directional output range reduction and (4) a direction-selective alternating current (AC) component. The after-potential and the AC component have a global effect in adapted neurons, by transferring to previously unstimulated parts of the receptive field (Nordström and O'Carroll, 2009). The contrast gain reduction and the output range reduction, however, are local and do not transfer to previously unstimulated parts of a neuron's receptive field.
Several recent studies have investigated the origin of motion adaptation, the dependence of intracellular calcium levels on adaptation, and to what extent adaptation improves information transmission (e.g. Clifford and Ibbotson, 2002; Fairhall et al., 2001; Kalb et al., 2008; Kurtz et al., 2000; Safran et al., 2007; Solomon et al., 2004). Although many of these studies look at the effects following several seconds of adaptation, more recent experiments using dynamic, fluctuating visual stimuli suggest that adaptation operates on the time scale of the core response of the elementary motion detector (EMD), and thus is likely to be an integral part of it (Borst et al., 2005; Wark et al., 2007). Furthermore, recent work from our own laboratory has highlighted the role of rapid motion adaptation in reliable encoding of the velocity of natural images (Barnett et al., 2010; Straw et al., 2008). However, it is not clear from previous work which motion adaptation components are involved in the fast and dynamic response changes. In this paper, we use a test–adapt–test protocol, which has previously been shown to be a powerful tool for dissociating the key components of adaptation (Harris et al., 2000). By varying the duration of the adapter over a large range, from 20 ms to 20 s, we are able to quantify, for the first time, the time course of each component. Our results underscore the speed at which adaptation operates, suggesting that under free flight the neurons are continuously adapting to the changing scenery. This might help explain recent observations of strong invariance in the response to natural scenes with highly variable contrast and image structure (Barnett et al., 2010; Straw et al., 2008).
MATERIALS AND METHODS
Hoverflies (Eristalis sp. Latreille 1804) were caught under permit and stored at 4°C. Prior to experiments they were waxed down with the head tilted forward to gain access to the posterior head surface. We cut a small hole over the left lobula complex and recorded intracellularly from horizontal system (HS) neurons using aluminosilicate electrodes pulled on a Sutter P-97 electrode puller (Novato, CA, USA). The electrodes were filled with 2 mol l–1 KCl and had a typical tip resistance of 50–150 MΩ. All experiments were carried out at 24–28°C.
The flies were mounted in front of an RGB CRT monitor with a refresh rate of 200 Hz, a mean luminance of 100 Cd m–2 and a spatial resolution of 640×480 pixels. We displayed full-screen sinusoidal gratings using a test–adapt–test protocol [adapted from Harris et al. (Harris et al., 2000)]. During 1 s test phases the grating moved in the preferred direction at a near-optimal combination of spatial and temporal frequency (0.1 cycles deg–1, 5 Hz). We then varied the contrast of the test grating to determine the full contrast response function, before and after adaptation. During the adapting phase, the grating moved in the anti-preferred direction for variable durations at a temporal frequency of 20 Hz. As our monitor had a refresh rate of 200 Hz, the shortest adapting duration we could confidently use as a motion stimulus was 20 ms (i.e. four frames). Because of time constraints using intracellular recordings, we used 20 s as the longest adapting duration. We included a minimum 2 s rest between the first test and the adapting stimulus, and a minimum 6 s rest between trials.
As control, we investigated the effect of adapting with flicker stimuli at 20 Hz displayed for 50 ms to 1 s. At 20 Hz, 50 ms corresponds to one full flicker cycle. We adapted with whole-field flicker, or with counter-phase grating flicker (0.1 cycles deg–1), as indicated in the text.
To quantify the after-potential we measured the response to a blank screen of mean luminance, following adaptation. By subtracting the after-potential from the adapted response, we normalized the adapted response (Harris et al., 2000). To quantify the output range reduction, we divided the normalized response to a test contrast of 1.0 by the unadapted response to the same test contrast. To quantify the AC component, we performed a fast Fourier transform (FFT) of the raw data (before spikelet removal) and quantified the mean power spectral density between 70 and 110 Hz (see Nordström and O'Carroll, 2009). All these analyses were performed within a 100–300 ms time window post test-stimulus onset.
To identify the earliest possible time window for response analysis, we defined a threshold based on the standard deviation of the response fluctuation in the unstimulated neuron (the mean standard deviation of the membrane fluctuation for 2 s before the start of test 1). We then identified the time point at which the averaged response (N=19) to a test contrast of 1 passed the threshold, as a measure of absolute response latency. We tested in both the preferred and anti-preferred direction and subtracted the responses from each other (i.e. PNP–NNN, where P is preferred and N is null) to avoid measuring the bleed-over response to the adaptor. We quantified the adapted and unadapted responses for 50 and 100 ms immediately following the absolute response latency, for each individual neuron, as above. We normalized the subtracted data to the gain (Eqn 1) of the unadapted contrast response function.
We analyzed all data statistically using GraphPad Prism software (La Jolla, CA, USA), with two-way ANOVAs, followed by paired t-tests (as indicated) with significance allocated to P<0.05 (after Bonferroni correction, where necessary). All data are displayed as means ± s.e.m., unless otherwise stated, where N represents the number of neurons and n the total number of trials.
To investigate the effect of different adapting durations, we used a test–adapt–test protocol (Fig. 1) where we quantified the contrast response function to full-screen sinusoidal gratings before and after adapting hoverfly HS neurons in the anti-preferred direction, thus allowing us to separate the different components of motion adaptation (Fig. 1D) (see Harris et al., 2000; Nordström and O'Carroll, 2009). HS neurons respond to high-contrast gratings in a direction-selective manner, depolarizing in the preferred direction (P, tests 1 and 2; Fig. 1A,B) and hyperpolarizing to motion in the opposite direction (N, adaptation). Small depolarizing monophasic events, called spikelets, ride on top of the graded response. The role of spikelets in neural coding is not completely established, but they likely increase the dynamic range of information transmission (Beckers et al., 2007; Beckers et al., 2009; Haag and Borst, 1996; Haag and Borst, 1998). As spikelets are monophasic depolarizing events (Haag and Borst, 1996), they boost the apparent membrane potential. As in our previous study (Nordström and O'Carroll, 2009), we therefore determined the underlying generator potential by removing the spikelets before averaging responses (black trace in Fig. 1B). As previously (Harris et al., 2000; Nordström and O'Carroll, 2009), we measured the response between 100 and 300 ms following stimulus onset (boxed in Fig. 1B).
Four components of motion adaptation
Contrast gain reduction
Following prolonged adaptation in either direction, motion-sensitive neurons showed a pronounced contrast gain reduction (Harris et al., 2000; Kohn and Movshon, 2003) evident by a rightward shift of the contrast response function in the adapted neuron (Fig. 1D). Fig. 2 shows the full contrast response functions before (filled symbols) and after adaptation in the anti-preferred direction (open symbols) for durations ranging from 20 ms to 20 s. Interestingly, we see evidence of substantial contrast gain reduction following even very brief adapting durations (Fig. 2B–D). The unadapted responses represent the response to test 1 for each trial, and thus serve as an internal control to ensure that the adaptation level of the neuron is reset between trials. The contrast gain reduction after the shortest adapting durations is not a residual response to the first test grating, as the unadapted and adapted contrast sensitivity functions overlie each other perfectly in the control, where the adapting duration was 0 s (Fig. 2A).
By subtracting the after-potential (measured as the adapted response to a zero contrast test pattern, i.e. a blank mid-luminance screen; see Fig. 1C,D) we normalized the data (dashed lines in Fig. 2). The after-potential was relatively small after adapting in the anti-preferred direction [see e.g. Harris et al. (Harris et al., 2000), Kurtz (Kurtz, 2007) and Nordström and O'Carroll (Nordström and O'Carroll, 2009) for longer adapting durations] and the normalized data also show a substantial contrast gain reduction after these short adapting durations (Fig. 2B–D). We quantified the contrast gain reduction by measuring C50 (Fig. 3A), defined as the contrast needed to generate a 50% maximum response in the unadapted neuron after fitting a Weibull function to the data. The contrast gain reduction increased with adapting duration (Fig. 2E–H, Fig. 3A), reaching a plateau after adapting for several seconds (Fig. 2I–L, Fig. 3A). Increasing the adapting duration further therefore seemed to have little additional effect on the contrast gain reduction. The adapted responses were significantly different from the unadapted responses for each adapting duration (paired t-tests, P<0.05; Fig. 3A).
The antagonistic after-potential
Following several seconds of stimulation, visual neurons displayed an antagonistic (direction-selective) after-potential, also called a motion after-effect, which lasted longer and was more pronounced following motion in the preferred direction (Harris et al., 2000; Kurtz et al., 2000). When we used a test contrast of 0 (Fig. 1C) the after-potential was evident in the raw data following 1 s of anti-preferred direction adaptation. As it is direction selective, the after-potential is manifested as a vertical shift in either direction (up or down) in the contrast response function (Fig. 1D).
To determine the effect of the adapting duration on the after-potential, we measured the adapted response to the zero-contrast test stimulus following each adapting duration (Fig. 1C). Our data (Fig. 3B) show a small antagonistic after-potential after ≥500 ms of anti-preferred adaptation, which remains at a fairly constant level with increasing durations [0.5–20 s, mean 1.37 mV; Fig. 3B; consistent with our earlier work (Harris et al., 2000; Nordström and O'Carroll, 2009)]. There was no significant difference between the after-potential following adapting durations of 0.5–20 s, consistent with an earlier study that found no increased after-potential when using anti-preferred direction adaptation durations of 200 ms to 4 s (Kurtz et al., 2000). However, the after-potential following preferred direction adaptation increased with the adapting duration (Kurtz et al., 2000) (see also supplementary material Fig. S1). Other work (Harris et al., 2000) showed that the strength of the after-potential is proportional to the strength of the response during stimulation.
Output range reduction
Adaptation for several seconds in either direction is followed by an output range reduction (Harris et al., 2000; Kohn and Movshon, 2003), which generates a compression of the contrast sensitivity function to high-contrast stimuli (Fig. 1D). To determine the effect of the adapting duration on the output range reduction, we measured the responses to a test contrast of 1.0 and divided the normalized response (adapted – after-potential) by the unadapted response. As control we measured the output range reduction after 0 s adaptation, which predicatively has a value close to 1 (0 s adapting duration; Fig. 3C). The output range reduction reached a value significantly different than this control, following adaption for 500 ms or more (stars, Fig. 3C). The output range reduction reached a plateau after adapting for several seconds of adaptation, with longer durations having no additional effect.
We recently described a new global component of motion adaptation, which is most prominent following several seconds of anti-preferred direction motion (Nordström and O'Carroll, 2009). The AC component boosts adapted responses to low-contrast stimuli by increasing the frequency of spikelets generated in the immediate post-adapting period. The fast sodium currents that cause spikelets lead to an amplification at approximately 100 Hz (Haag and Borst, 1996).
To quantify the AC component after different adapting durations, we measured the power of the response at 70–110 Hz after performing a FFT of the raw data (i.e. before removing spikelets) within the same analysis window as before (Fig. 3D). The AC power varies between neurons, and is susceptible to small drifts in the resting membrane potential. In a closer analysis of the individual responses, we noted that two neurons act as outliers (>2×s.d.) following most adapting durations (gray triangles, Fig. 3D). The data for these two neurons were therefore excluded before calculating the means (circles, Fig. 3D). For the remaining neurons, the AC component increased with longer adapting durations.
Adapting with flicker
HS neurons, like other LPTCs, give a strong transient flicker response at the onset and offset of high-contrast visual stimuli (e.g. Borst and Egelhaaf, 1987; Nordström and O'Carroll, 2009). Non-directional flicker stimuli also resulted in reduced contrast sensitivity (i.e. contrast gain reduction) (Harris et al., 2000). Although this effect was weaker than the motion-induced contrast gain reduction, it is possible that flicker-dependent gain reduction represents a contribution of contrast adaptation in earlier stages of visual processing (e.g. in the lamina), which may occur on a different time scale to motion-specific adaptation. To investigate whether the effects following short adapting durations (Fig. 3) may represent a flicker response to the adapting stimulus, we replaced the moving adaptor with a high-contrast, whole-field 20 Hz flicker (1 s adaptation, see Fig. 4A,B). The flicker thus stimulates the eye at the same frequency as the motion adaptor, but with no relative motion cues. During such flicker adaptation, the HS membrane potential is modulated around the resting level (Fig. 4B).
The full contrast sensitivity functions after adapting with flicker at five different durations (50 ms to 1 s) are shown in Fig. 4C–G. Following even the shortest adapting duration (50 ms, Fig. 4C), we see a rightward shift of the curve (contrast gain reduction), which becomes more pronounced after longer adapting durations (Fig. 4F,G). As earlier, we quantified the contrast gain reduction by determining C50 in the unadapted and adapted neuron after fitting the data with a Weibull function, and found that flicker induces a significant contrast gain reduction following ≥100 ms adaptation (hashes, Fig. 4H). Earlier work (Harris et al., 2000) showed C50 values that were ca. 1.5 times larger following 4 s motion adaptation compared with flicker adaptation, which is consistent with our data (the normalized C50 values following anti-preferred motion adaptation are on average 1.7 times larger than following flicker; Fig. 4H; motion data are re-plotted from Fig. 3A for direct comparison).
There is no after-potential after adapting with whole-field flicker (squares, Fig. 4I), consistent with earlier work using 4 s of flicker adaptation (Harris et al., 2000). This differs from adapting with motion, which induces an after-potential after ≥500 ms anti-preferred adaptation (circles, Fig. 4I; data re-plotted from Fig. 3B).
The output range is not reduced as much following flicker adaptation (average 0.94) as after motion (average 0.91) for adapting durations between 50 ms and 1 s, and is never significantly different from the 0 s adaptation condition (Fig. 4J), consistent with our previous work showing that 4 s of flicker adaptation generates a negligible output range reduction (Harris et al., 2000). Our data (Fig. 4J), and the finding that the output range reduction is more pronounced after adapting with orthogonal motion than with flicker (Harris et al., 2000), suggest that this component is generated after the computation of motion.
There is no consistent difference in the post-adapted AC component following flicker compared with motion adaptation (Fig. 4K). Following flicker adaptation, the mean power spectral density is significantly increased compared with the unadapted response after 100 and 200 ms adaptation, but as there is no consistent trend with increasing adapting durations it is hard to rule out noise in the data.
Early contrast gain reduction
Of the four components of motion adaptation investigated here, the contrast gain reduction appears earliest (Fig. 3). Previous work showed that adaptation operates on varying time scales ranging from tens of milliseconds to several minutes, and that neurons adapt and unadapt on similar time scales (Fairhall et al., 2001; Wark et al., 2007). This suggests that in order to measure the effect following very short adapting durations (e.g. 20 ms) we need to look at the post-adapted response within similar time scales, or the effect might already have disappeared. This is a difficult analysis, however, because the neurons display an absolute response latency of several tens of milliseconds. Therefore, we need to determine the absolute response delay under our experimental conditions. As shown in previous work (Barnett et al., 2010; Warzecha and Egelhaaf, 2000), the response latency increases with decreasing contrast, and the shape of the initial response is highly dependent on the stimulus contrast – whereas high-contrast stimuli display a sharp start-up transient, lower-contrast stimuli build up over several hundred milliseconds (mean responses to test 1 of five different contrasts, N=19, n=240; Fig. 5A).
We identified a threshold response based on the standard deviation of the membrane potential fluctuation in the unstimulated neuron (thin lines, Fig. 5B). In the magnified inset in Fig. 5A, we see that the averaged response to a test contrast of 1.0 passes the standard deviation (i.e. our threshold, thin line in Fig. 5A) 28.3 ms after stimulus onset. The earliest time we can measure the adapted response is thus after ca. 30 ms. However, 30 ms post test-stimulus onset, the generator potential may reflect the response to a brief adapting stimulus, which would also take approximately 30 ms to pass the threshold, rather than reflect the response to the test stimulus currently displayed. To avoid the bleed-over contribution of the response to the adaptor, we subtracted preferred responses from non-preferred direction responses following the same type of adaptation (PPP–NPN or PNP–NNN; Fig. 5B) (and see Barnett et al., 2010). Following such a subtraction, there is still a substantial contrast gain reduction present immediately following the absolute response delay (30–80 ms post-stimulus onset; Fig. 5C). We determined C50 as before to quantify the contrast gain reduction (Fig. 5F). As low-contrast stimuli will not have reached the threshold in this 50 ms time window (Fig. 5A), we increased the analysis to 100 ms (30–130 ms post-stimulus onset), and again, found a substantial contrast gain reduction (Fig. 5G). This analysis thus confirms the data shown in Fig. 3A by supporting the presence of motion-induced contrast gain reduction after as little as 20 ms adaptation.
As the contrast gain reduction did not saturate with longer flickering adapting durations and was also weaker after flicker than after motion at all durations (Fig. 4H), it is unlikely to be caused by contrast adaptation in early vision. It could potentially be an inherent effect of the EMD, as flicker stimuli generate a non-directional oscillation due to imbalance in the subtraction mechanism, leading to a weak sensitivity to flicker (Egelhaaf et al., 1989). To investigate the degree to which the very rapid contrast gain reduction (Fig. 5C,F) is an effect of such inherent flicker sensitivity of the EMD, we performed a similar subtraction after adapting with counter-phase flicker (i.e. PFP–NFN) instead of motion (Fig. 5D). As above, this allowed us to tease apart the response to the test stimulus from the bleed-over contribution of the response to the flicker adaptor. The counter-phase flicker stimulus stimulates the underlying EMDs with the same spatial and temporal frequency as the motion adaptor, without providing relative motion cues. Because of spatial lateral inhibition at the inputs to the EMDs, a counter-phase flicker may act as a stronger adaptor than a full-field flicker.
Following such a subtraction, we found that flicker adaptation generated much less contrast gain reduction than motion (Fig. 5D,H). With longer adapting durations, the contrast gain reduction increased (Fig. 5H), consistent with the data shown in Fig. 4H. It is thus unlikely that the rapid contrast gain reduction (Fig. 5C,F,G) is caused by flicker sensitivity generated by the subtraction stage of the EMD; rather, it is more likely to be caused by other mechanisms involved in the computation of motion.
To confirm the conclusion that the rapid contrast gain reduction is not a consequence of intrinsic transient properties of the EMD, we analyzed the output of a mildly elaborated EMD model based on a basic Hassenstein–Reichardt correlator with soft saturation on the output, and mildly elaborated inputs to account for spatiotemporal filtering by the lamina (see Materials and methods). There was no contrast gain reduction after any adapting duration (50 ms adaptation; Fig. 5E), but rather a slight leftward shift of the curve for the shortest adapting durations. In other words, interactions due to the transient nature of the basic EMD led to a weak facilitation of the adapted responses for the shortest adapting durations. Indeed, the possible facilitation resulting from transient interactions within the EMD means that our experiment, if anything, underestimates the reduction in contrast gain for these brief adapting stimuli.
Importantly, if the contrast gain reduction operates on a time scale of tens of milliseconds, this suggests that the unadapted responses measured in Figs 2–3 and in previous work (e.g. Harris et al., 2000; Kurtz et al., 2000) are already themselves adapted. To assess the responses earlier than the 100–300 ms window typically used, however, we have to use a subtraction method as described here (Fig. 5) to avoid measuring the response to the adaptor rather than to the test stimulus. This quickly becomes a prohibitively large experiment, and further, only allows for quantification of the non-directional components, i.e. the contrast gain reduction and the output range reduction. However, from such analysis we see no output range reduction after the shortest adapting durations (Fig. 5C), suggesting that the data for the two non-directional components, shown in Fig. 3, give a fair representation of the underlying adaptation.
Few previous studies have looked at the effect of varying the adapting duration. Although some (Kurtz et al., 2000) looked at the strength of the after-potential and the calcium concentration after different adapting durations, they did not investigate the other components of motion adaptation. The formation of after-images and the effects following adaptation with dynamic white noise stimuli have been investigated (Fairhall et al., 2001; Maddess, 1986); however, in these cases the different components of motion adaptation were not dissociated. As very large changes in sensitivity can be masked by saturated responses (e.g. to high-contrast stimuli), quantifying the contrast sensitivity function using a test–adapt–test approach is a powerful method for dissecting the time course of adaptation. To our knowledge this is the first time the four different components of adaptation (Table 1) have been thoroughly quantified following short adapting durations. We show that although contrast gain reduction is extremely rapid, the other three components evolve with similar time scales. All four components saturate after several seconds of stimulation. We also show that flicker sensitivity or other transient properties of the EMD are unlikely to explain the very rapid contrast normalization.
Contrast gain reduction takes place incredibly fast. The effect is non-directional and thus present after adaptation in either the preferred or anti-preferred direction (Harris et al., 2000). The contrast gain reduction is local, in that only those input dendrites directly stimulated by visual motion adapt (Nordström and O'Carroll, 2009), suggesting that it originates pre-synaptic to LPTCs. LPTCs spatially pool information from a large array of EMDs. These input elements to the LPTCs show only weak direction tuning, and the input dendrites serve to both spatially integrate the signal and enhance the direction selectivity (Borst and Haag, 2002; Single et al., 1997). It is possible that the contrast gain reduction is associated with decreased gain at the synaptic input to the LPTCs (see also Nordström and O'Carroll, 2009). Activity-dependent calcium accumulation in LPTCs has previously been linked to some components of adaptation (Dürr and Egelhaaf, 1999; Kurtz, 2007; Kurtz et al., 2000), but this is a relatively slow phenomenon, suggesting that other mechanisms may be responsible for the fastest changes seen here and in previous experiments with dynamic stimuli (Borst et al., 2005; Wark et al., 2007). The other local non-directional component, the output range reduction, is likely caused by other cellular mechanisms than the contrast gain reduction, as it is much slower (Fig. 3C, Table 1).
We have shown a potent contrast gain reduction on a time scale at least as fast, if not faster, than the 35 ms time constant of the low-pass filter of the EMD, i.e. adaptation is intrinsic to the detection of motion itself. This very rapid contrast normalization (extremely rapid motion-dependent adaptation) on time scales on par with the delay of the EMD provides a partial explanation for the observation that LPTCs display ‘velocity constancy’ (Straw et al., 2008), whereby the velocity tuning to a range of natural images is very similar despite the images having a large spread of contrasts. This observation has been difficult to reconcile with current models for motion estimation (Barnett et al., 2010; Straw et al., 2008), but if the neurons continuously adapt to the current local contrast on time scales shorter than the low-pass filter of the EMD, this could act to rapidly normalize responses across a large range of images. Using dynamic motion stimulation, with velocities that fluctuate rapidly (albeit using white noise rather than natural images), it has been shown that adaptation in LPTCs operates on the same time scale as the response itself (Borst et al., 2005). With increasing velocity fluctuations, the contribution to the response from previous times are suppressed (Borst et al., 2005; Fairhall et al., 2001) so that the system is continuously updating itself to the recent stimulus history.
In fact, during flight, insects are rarely exposed to continuous motion at a single velocity. Rather, they are exposed to a much more dynamic visual stimulus with a large range of spatial frequencies (Tolhurst et al., 1992) and velocities that vary over time (Boeddeker et al., 2005). Measurements of vertebrate eye movement dynamics suggest that these are matched to the substantial variation in local scenes, so that local adaptation builds up and decays rapidly within the time scale of a single fixation (Frazor and Geisler, 2006). Although an efficient coding strategy is to represent those inputs that a neuron typically processes, this becomes difficult when coding natural stimuli that differ greatly on a local scale (both in time and/or space) (Tolhurst et al., 1992) compared with their global distribution. This makes it hard to generate reliable coding, and rapid adaptation dependent on recent stimulus history becomes a prerequisite (Wark et al., 2007). The rapid adaptation we have shown here would no doubt play an important role in this.
LIST OF ABBREVIATIONS
elementary motion detector
lamina monopolar cell
lobula plate tangential cell
null test, flicker adapt, null test
null test, null adapt, null test
null test, preferred adapt, null test
preferred test, flicker adapt, preferred test
preferred test, null adapt, preferred test
preferred test, preferred adapt, preferred test
This work was funded by the Australian Research Council [grant number DP0880983], US Air Force Office of Scientific Research [grant number FA2386-10-1-4114], Swedish Research Council [grant number 2008-2933], The Royal Swedish Academy of Sciences [grant number FOA10H-119], and P. O. Zetterling Foundation.
We thank the Managers of the Botanic Gardens of Adelaide for allowing us to collect insects and for their ongoing support. We also thank the reviewers whose comments and suggestions greatly improved the manuscript.