The study of decision-making has largely been influenced by findings on saccadic decisions made by monkeys in a visual forced choice task with two alternatives (e.g., Gold and Shadlen 2007; Churchland et al. 2008; Hanks et al. 2014). In the experiments, a monkey views a screen from fixed distance and orientation and sees an assembly of dots, each of which can move either to the left or to the right. The monkey’s task is to decide if more dots move to the left or to the right. The task is made harder by decreasing the percentage of dots moving in one direction from 100 to 50%. From these studies, it has become clear that (1) the decisions become less accurate when they must be made in lesser time, (2) when the animal is free to refuse cooperating in the task it will do so when accuracy is low, and (3) raising the number of options from two to four decreases the accuracy of the decisions (Churchland et al. 2008). The findings not only capture many aspects of decision-making but could be explained simply by recordings in the lateral intraparietal cortex. With increasing evidence in favor of each of the two options, two sets of neurons, one for each option, increased their firing rate until one of them first reaches a threshold. The success of this model should, however, not make us forget that there are decisions that simply do not offer the time for such a solution and yet need to be complex and accurate. The archerfish’s predictive C-start decision is a good example for this. It fails to obey any of the characteristics (1)–(3) noted above: Accuracy is unrelated both to latency and to the probability that the fish will respond with a predictive start. Furthermore, even adding a new variable (adding many more than two additional options) did not affect latency or accuracy (Reinel and Schuster 2016). This is not because the predictive start decisions are simple and made between a limited number of pre-set options (Wang 2008). To the contrary, these decisions are made in a continuum of four variables: vertical initial speed, horizontal initial speed, azimuthal direction, and initial height. They are not made from a fixed vantage point but allow the fish to achieve accuracy in speed and turn angle over the full range of orientations and from all relevant distances. It should be stimulating to understand why some complex decisions (complex by the large number of options) apparently are not bound by the type of constraints that are seen in the saccadic decisions of monkeys. In rapid decisions, such as the archerfish’s start decisions, other mechanisms must have evolved to combine utmost speed with accuracy.

To quickly bend into the shape of a letter ‘C’, it is essential that all trunk muscles contract on only one side, and for maximum force all muscle fibers should contract simultaneously. This can be achieved by sending the command down a thick and fast-conducting axon. The Mauthner neuron is ideally suited for just this task (e.g., Korn and Faber 2005; Sillar 2009; Sillar et al. 2016). Each axon of the two Mauthner neurons crosses the fish’s midline and runs down the spinal cord. The two neurons are wired such that only one of them fires one and only one action potential. So, if the Mauthner neuron located on the left side of the fish gets activated, it sends an action potential down the right side of the fish (because the axon crosses the midline), whose rapid spreading causes almost simultaneous contraction of trunk muscles on the right side, thus powerfully bending the fish towards the right. As simple as this view is it has met with remarkably fierce resistance. First, ablation experiments have repeatedly found at least some equally short-latency, high-power C-starts (e.g. DiDomenico et al. 1988; Eaton et al. 1982, 1991, 1995; Gahtan and Baier 2004; Lacoste et al. 2015; Liu and Fetcho 1999; Zottoli et al. 1999). Second, it has been argued that a few milliseconds of increase in latency would not be of any consequence for survival (DiDomenico et al. 1988; DiDomenico and Eaton 1988; Eaton et al. 1991, 1995). Moreover, even an unusually low capacity to regenerate was described particularly for the large Mauthner axon (Bhatt et al. 2004), making it even more cumbersome to assume why this neuron should be of particular relevance for triggering live-saving C-starts. The finding that low-latency high-power C-starts should still be possible when the Mauthner cells are lost does fit the generally accepted idea that no single neuron can underlie important behavioral functions (the textbook ‘grandmother neuron doctrine’). However, it makes it puzzling, why then these large neurons are still present, when their function can, at least in principle, be taken over by smaller neurons. It recently turned out, that completely unilaterally removing specifically one Mauthner neuron, including its axon, specifically removed the ability of zebrafish larvae to produce short-latency high-power C-starts to the side in which the axon was missing but not to the side in which it was still present (Hecker et al. 2020a). In these experiments, the complete and slow Wallerian degeneration of the axon was followed in a two-photon microscope until the complete axon was absent. Stimuli were given at various stages of the axon degeneration to probe C-starts. These stimuli were selected for their ability to elicit activity in all neurons of the so-called Mauthner series (e.g., Liu and Fetcho 1999), so that other neurons could in principle substitute for the absent Mauthner neuron. However, removal of the Mauthner axon completely removed all short-latency, high-power C-starts. This finding also allowed what appears to be the first direct experimental test of the debated issue of the actual survival value of these maneuvers. In tests caried out with a natural predator, sham-ablated larvae with intact Mauthner cells had a higher chance of surviving the predator attacks (Hecker et al. 2020a). Furthermore, raising unilaterally ablated larvae and examining them double-blind months later as adults showed that this deficiency is never restored even during the massive transformation from the larva to an adult fish (Hecker et al. 2020a). In other words, the Mauthner neuron is essential for driving powerful short-latency C-starts both in larval and adult fish and removing it does have clear consequences (Fig. 6a). It also turned out that—in line with the unique importance of this neuron—the regenerative capability of the Mauthner axon is not generally low but instead strongly dependent on where on its length the axon is lesioned. In the front end of the fish, where the effect of losing the axon on C-start latency is expected to be greatest, regeneration is impressively rapid (Hecker et al. 2020b).

Evidence used in the discussion of an involvement of the archerfish Mauthner neuron (MN) in triggering archerfish C-starts. a The importance of the MN in adult fish. Larval zebrafish had either their right or left MN (including its axon) ablated as confirmed using two-photon microscopy. They were subsequently raised together with untreated siblings. At least 5 months later C-starts were elicited by a standardized stimulus that caused the fish to bend to the left or to right side. After accumulating enough starts to both sides, each experimental fish was sacrificed to determine whether the left or the right Mauthner axon was missing (or none). Diagrams show the findings for the unilaterally ablated fish and demonstrate that, in the same individuals (lines in left diagram), latency was always shorted in C-starts that could recruit the remaining Mauthner axon (ipsi) than in the starts that could not (contra). CDF = cumulative distribution function. b Intracellular filling of the MN in equally sized goldfish (red) and archerfish (blue) revealed no major differences. See Machnik et al. (2018a) for detailed measurements. c Left: The axons of the archerfish MN are by far the largest in the spinal cord, but their diameter is not significantly different from that in goldfish of similar total length (right). d The only difference found in a detailed comparative analysis was that in archerfish postsynaptic potentials (PSPs) elicited by acoustic stimuli and light flashes, whose intensities were chosen to elicit PSPs of comparable size in goldfish, were larger for the visual stimuli. Adapted from Hecker et al. (2020a), Machnik et al. (2018a, b)

Archerfish do have Mauthner neurons (Fig. 6b) that have a so-called axon cap, which means that they can be found in the hindbrain by an electrical signature, the same way as the goldfish Mauthner cell is found (Machnik et al. 2018a, b). The archerfish Mauthner axon has the largest diameter of all axons that run down the archerfish spinal cord (Fig. 6c). Both morphologically and functionally the archerfish Mauthner neurons are like those of the goldfish with the only major difference being their higher visual sensitivity. So, the key question now is whether archerfish can use an apparently ‘standard’ Mauthner neuron for something as sophisticated as their predictive C-start decisions. An easily acceptable view would seem that archerfish use this neuron for escapes but not for their predictive starts. But why then are the escape C-starts and the predictive C-starts not easily distinguishable from their kinematics? Is there perhaps some other circuitry that achieves the same top-performance as the escape C-start? If that was the case, then, again, the use of other (smaller) neurons would allow to produce top-performance C-starts so that similar networks of smaller cells with thinner axons could also be used to produce top-performance escape C-starts, making the archerfish Mauthner neuron obsolete. The simplest explanation, also considering the findings in adult zebrafish (Fig. 6a), is that the archerfish C-starts, escapes and predictive starts alike, are initiated by an action potential that travels down the thick axon of one of the Mauthner neurons. Clearly, a remarkable and highly efficient preprocessing is needed, and this will probably differ among escapes and predictive starts, but the final ‘go’ will have to be given by firing one of the two Mauthner neurons. The archerfish predictive C-starts, thus, lead to the intriguing question of what can and cannot be done with an individual heavily compartmentalized (e.g., Korn and Faber 2005) identified neuron in a vertebrate brain.