Author response:
Public Reviews:
Reviewer #1 (Public review):
Summary:
The authors test whether the archerfish can modulate the fast response to a falling target.
We have not tested whether archerfish can 'modulate the fast response'. We quantitatively test specific hypotheses on the rules used by the fish. For this the accuracy of the decisions is analyzed with respect to specific points that can be calculated precisely in each experiment. The ill-defined term 'modulate' does in no way capture what is done here. This assessment might explain the question, raised by the reviewer, of 'what is the difference of this study and Reinel, 2016' (i.e. Reinel and Schuster, 2016). In that study, all objects were strictly falling ballistically, and latency and accuracy of the turn decisions were determined when the initial motion was not only horizontal but had an additional vertical component of speed. The question of that study was if the need to account to an additional variable (vertical speed) in the decision would affect its latency or accuracy. The study showed that also then archerfish rapidly turn to the later impact point. It also showed that accuracy and latency (defined in this study exactly as in the present study) were not changed by the added degree of freedom. This is a completely different question and by its very nature does not leave the realm of ballistics.
By manipulating the trajectory of the target, they claim
that the fish can modulate the fast response.
While it is clear from the result that the fish can modulate the fast response, the experimental support for the argument that the fish can do it for a reflex-like behavior is inadequate.
This is disturbing: The manuscript is full of data that directly report response latency (a parameter that's critical in all experiments) and there are even graphical displays of the distribution of latency (Figs. 2, 5). How fast the responses are, can also already be seen in the first video. Most importantly, the nature of the 40 ms limit has been discovered and has been reported by our group in 2008 (Schlegel and Schuster, 2008, Fig. 4). For easy reference, we attach Schlegel and Schuster, 2008 with the relevant passages marked in yellow. But later studies also using high speed video (ie. typically 500 fps) and simultaneously evaluating accuracy and kinematics (in the same ways as used here!) to address various questions repeatedly report and even graphically represent minimum latencies of 40 ms, e.g. Krupczynski and Schuster, 2013 (e.g. Fig. 2); Reinel and Schuster, 2014; Reinel and Schuster, 2016; Reinel and Schuster, 2018a, b (e.g. see Fig. 7 in the first part) and report how latency is increased as urgency is decreased (if the fish are too close or time of falling is increased), as temperature is decreased or as viewing conditions and their homogeneity across the tank change. Moreover, even a field study is available (Rischawy, Blum and Schuster, 2015) that shows why the speed is needed. This is because of massive competition with at least some of the competitor fish also be able to turn to the later impact point. So, speed is an absolute necessity if competitors are around. Interestingly, when the fish are isolated, latency goes up and eventually the fish will no longer respond with C-starts (Schlegel and Schuster, 2008).
Another aspect: considering the introduction it would not even have mattered if not 40 ms but instead 150 ms were the time needed for an accurate start (which is not the case). That would still be faster than an Olympic sprinter responds to a gun shot. Moreoever, please also note that we carefully talk of reflex-speed not of a reflex-behavior (which is as easy to verify as any other if the false statements made).
Strengths:
Overall, the question that the authors raised in the manuscript is interesting.
Given the statement of no difference between the present study and Reinel and Schuster, 2016, it is not clear what this assessment refers to.
Weaknesses:
(1) The argument that the fish can modulate reflex-like behavior relies on the claim that the archerfish makes the decision in 40 ms. There is little support for the 40 ms reaction time.
The 'little support' is a paper in Science in which this important aspect is directly analyzed (Fig. 4 of that paper) and that has been praised by folks like Yadin Dudai (e.g . in Faculty 1000). The support is also data on latency as presented in the present paper. Furthermore, additional publications are available on the reaction time (see above).
The reaction time for the same behavior in Schlegel 2008, is 60-70 ms, and in Tsvilling 2012 about 75 ms, if we take the half height of the maximum as the estimated reaction time in both cases. If we take the peak (or average) of the distribution as an estimation of reaction time, the reaction time is even longer. This number is critical for the analysis the authors perform since if the reaction time is longer, maybe this is not a reflex as claimed.
See above.
In addition, mentioning the 40 ms in the abstract is overselling the result.
See above.
Just for completeness: Considering a very interesting point raised by reviewer 2 we add an additional panel to further emphasize the exciting point that accuracy and latency are unrelated in the start decisions. That point was already made in Fig.4 of the paper in Science but can be directly addressed.
The title is also not supported by the results.
No: the title is clearly supported by the results that are reported in the paper.
(2) A critical technical issue of the stimulus delivery is not clear.
The stimulus delivery is described in detail. Most importantly we emphasize (even mentioning frame rate) that all VR setups require experimental confirmation that they work for the species and for the behavior at hand. Ideally, they should elicit the same behavior (in all aspects) as a real stimulus does that the VR approach intends to mimic. Whether VR works in a given animal and for the behavior at hand in that animal cannot be known or postulated a priori. It must be shown in direct critical experiments. Such experiments and the need to perform them are described in detail in Figure 2 and in the text that is associated with that figure.
The frame rate is 120 FPS and the target horizontal speed can be up to 1.775 m/s. This produces a target jumping on the screen 15 mm in each frame. This is not a continuous motion. Thus, the similarity between the natural system where the target experiences ballistic trajectory and the experiment here is not clear. Ideally, another type of stimulus delivery system is needed for a project of this kind that requires fast-moving targets (e.g. Reiser, J. Neurosci.Meth. 2008).
See above. It is quite funny that one of the authors of the present study had been involved in developing a setup with a complete panorama of 6000 LEDs (Strauss, Schuster and Götz, 1997; and appropriately cited in Reiser) that has been the basis for Reiser. This panorama was also used to successfully implement VR in freely walking Drosophila (Schuster et al., Curr. Biol., 2002). However, an LED based approach was abandoned because of insufficient spatial resolution (that, in archerfish, is very different from that of Drosophila).
But the crucial point really is this: Just looking at Figure 2 shows that our approach could not have worked better in any way - it provided the input needed to cause turn decisions that are in all aspects just as those with real objects. Achieving this was not at all trivial and required enormous effort and many failed attempts. But it allows addressing our questions for the first time after 20 years of studying these interesting decisions.
In addition, the screen is rectangular and not circular, so in some directions, the target vanishes earlier than others. It must produce a bias in the fish response but there is no analysis of this type.
Why 'must' it produce a bias? Is it not conceivable that you can only use a circular part of the screen? Briefly, and as could have been checked by quickly looking into the methods section, this is what we did. But still, why would it have mattered in our strictly randomized design? It could have mattered only in a completely silly way of running the experiments in which exclusively long trajectories are shown in one condition and exclusively short ones in another.
(3) The results here rely on the ability to measure the error of response in the case of a virtual experiment. It is not clear how this is done since the virtual target does not fall.
Well, of course it does not fall!!! That is the whole point that enables the study, and this is explained in connection with the glass plate experiment of Fig. 1 and quite some text is devoted to say that this is the starting point for the present analysis. The ballistic impact point is calculated (just as explained in our very first paper on the start decisions, Rossel, Corlija and Schuster, 2002) from the initial speed and height of the target, using simple high-school physics and the justification for that is also in that paper. This has been done already more than 20 years ago. How else could that paper have arrived at the conclusion that the fish turned to the virtual impact point even though nothing is falling? We also describe this for the readers of the present study, illustrate how accuracy is determined in Figures, in all videos and in an additional Supplementary Figure. Consulting the paper reveals that orientation of the fish is determined immediately at the end of stage 2 of its C-start and the error directly reports how close continuing in that direction would lead the fish to the (real or virtual) impact point. This measure has also been used since the first paper in 2002 in our lab and it is very useful because it provides an invariant measure that allows pooling all the different conditions (orientation and position of responding fish as well as direction, speed and height of target).
How do the authors validate that the fish indeed perceives the virtual target as the falling target?
See above. The fish produce C-starts (whose kinematics are analyzed and reported in Figures), whose latency is measured (from onset of target motion to onset of C-start) and whose accuracy in aligning them to the calculated virtual impact point is measured (see above). Additionally, the errors are also analyzed to other points of interest, for instance landmarks, the ballistic landing point in the re-trained fish or points calculated on the basis of specific hypotheses in the generalization experiments.
Since the deflection is at a later stage of the virtual trajectory, it is not clear what is the actual physics that governs the world of the experiment.
As explained in the text what we need is substituting the ballistic connection with another fixed relation between initial target motion and the landing point. This other relation needs to produce a large error in the aims when they remain based on the ballistic virtual landing point. It is directly shown in the key experiments that the fish need not see the deflection but can respond appropriately to the initial motion after training (Figs. 3, 5 and corresponding paragraphs in the text as well as additional movies). Please also note that after training the decision is based on the initial movement. This is shown in the interspersed experiments in which nothing than the initial (pre-deflection) movement was shown.
Overall, the experimental setup is not well designed.
It is obviously designed well enough to mimic the natural situation in every aspect needed (see Fig. 2) and well enough to answer the questions we have asked.
Reviewer #2 (Public review):
Summary:
This manuscript studies prey capture by archer fish, which observe the initial values of motion of aerial prey they made fall by spitting on them, and then rapidly turn to reach the ballistic landing point on the water surface. The question raised by the article is whether this incredibly fast decision-making process is hardwired and thus unmodifiable or can be adjusted by experience to follow a new rule, namely that the landing point is deflected from a certain amount of the expected ballistic landing point. The results show that the fish learn the new rule and use it afterward in a variety of novel situations that include height, side, and speed of the prey, and which preserve the speed of the fish's decision. Moreover, a remarkable finding presented in this work is the fact that fish that have learned to use the new rule can relearn to use the ballistic landing point for an object based on its shape (a triangle) while keeping simultaneously the 'deflected rule' for an object differing in shape (a disc); in other words, fish can master simultaneously two decision-making rules based on the different shape of objects.
Strengths:
The manuscript relies on a sophisticated and clever experimental design that allows changing the apparent landing point of a virtual prey using a virtual reality system. Several robust controls are provided to demonstrate the reliability and usefulness of the experimental setup.
Overall, I very much like the idea conveyed by the authors that even stimuli triggering apparently hardwired responses can be relearned in order to be associated with a different response, thus showing the impressive flexibility of circuits that are sometimes considered mediating pure reflexive responses.
Thank you so much for this precise assessment of what we have shown!
This is the case - as an additional example - of the main component of the Nasanov pheromone of bees (geraniol), which triggers immediate reflexive attraction and appetitive responses, and which can, nevertheless, be learned by bees in association with an electric shock so that bees end up exhibiting avoidance and the aversive response of sting extension to this odorant (1), which is a fully unnatural situation, and which shows that associative aversive learning is strong enough to override preprogrammed responding, thus reflecting an impressive behavioral flexibility.
That's very interesting, thanks.
Weaknesses:
As a general remark, there is some information that I missed and that is mandatory in the analysis of behavioral changes.
Firstly, the variability in the performances displayed. The authors mentioned that the results reported come from 6 fish (which is a low sample size). How were the individual performances in terms of consistency? Were all fish equally good in adjusting/learning the new rule? How did errors vary according to individual identity? It seems to me that this kind of information should be available as the authors reported that individual fish could be recognized and tracked (see lines 620-635) and is essential for appreciating the flexibility of the system under study.
Secondly, the speed of the learning process is not properly explained. Admittedly, fish learn in an impressive way the new rule and even two rules simultaneously; yet, how long did they need to achieve this? In the article, Figure 2 mentions that at least 6 training stages (each defined as a block of 60 evaluated turn decisions, which actually shows that the standard term 'Training Block' would be more appropriate) were required for the fish to learn the 'deflected rule'. While this means 360 trials (turning starts), I was left with the question of how long this process lasted. How many hours, days, and weeks were needed for the fish to learn? And as mentioned above, were all fish equally fast in learning? I would appreciate explaining this very important point because learning dynamics is relevant to understanding the flexibility of the system.
First, it is very important to keep the question in mind that we wanted to clarify: Does the system have the potential to re-tune the decisions to other non-ballistic relations between the input variables and the output? This would have been established if one fish was found capable of doing that. However, we do have sufficient evidence to say that all six fish learned the new law and that at least one (actually four) individual was capable of simultaneously handling the two laws. We will explain this much better (hopefully) in our revised version. We also have to stress that not all archerfish might actually be able to do this and that not all archerfish might learn in the same way, at the same speed, or using the same strategies. These questions are extremely interesting and we therefore definitely will include all evidence that we have. If some individuals are better than others in quickly adjusting, then even observational learning could become a part of the story. However, we needed to make and document the first steps. Understanding these is essential and apparently is difficult enough.
Reference:
(1) Roussel, E., Padie, S. & Giurfa, M. Aversive learning overcomes appetitive innate responding in honeybees. Anim Cogn 15, 135-141, doi:10.1007/s10071-011-0426-1 (2012).
Thanks for this reference!
