Neurons in nBOR and LM are ideally suited for the analysis of optic flow because they have large receptive fields averaging about 60° in diameter with some encompassing the entire monocular visual field. These neurons are directionally selective in response to large stimuli, such as random dot patterns, checkerboards, and gratings. The response properties of nBOR neurons have been investigated in chickens (Gallus gallus, Burns and Wallman, 1981) and pigeons (Morgan and Frost, 1981; Columba livia, Gioanni et al., 1984; Wylie and Frost, 1990), and the responses are essentially identical. With respect to direction preference, neurons that prefer upward, downward and backward (i.e., nasal-to-temporal) motion are about equally abundant in nBOR, but fewer (5–10%) prefer forward (i.e., temporal-to-nasal) motion (Frost et al., 1990). With respect to stimulus speed, neurons are broadly tuned, but the majority (~75%) prefer slow velocities (< 5°/s) (Burns and Wallman, 1981; Crowder et al., 2003). Data from chickens, pigeons and zebra finches (Taeniopygia guttata) suggest that the LM is complementary to the nBOR regarding direction preference. In LM, there is a clear directional bias to forward motion: about half the neurons prefer motion in this direction and neurons that prefer upward, downward and backward motion are equally represented (McKenna and Wallman, 1985b; Winterson and Brauth, 1985; Wylie and Frost, 1996; Wylie and Crowder, 2000; Gaede et al., 2017) (Figure 2D). The bias toward neurons that prefer forward motion has also been found in the homologs of the LM in amphibians, reptiles, and mammals (Collewijn, 1975; Katte and Hoffmann, 1980; Fan et al., 1995). With respect to speed, as in nBOR, LM neurons are broadly tuned, but most (about 65%) prefer faster stimuli (>5°/s and up to 250°/s). Almost all (>95%) of the LM neurons that prefer slower speeds prefer forward motion. Stated another way, there are two groups of LM neurons: (i) slow neurons that prefer forward motion; and (ii) fast neurons that prefer upward downward and backward motion (Wylie and Crowder, 2000) (see Figure 2F). There is also evidence suggesting that there is a separation of directional response types within LM. Winterson and Brauth (1985) noted that most cells in the LM parvocellularis, now known as the lateral LM (LMl; Gamlin and Cohen, 1988b) preferred forward motion, whereas cells that prefer upward, downward, backward and forward motion are found in LM magnocellularis, now known as the medial LM (LMm, Gamlin and Cohen, 1988b). This is important because the LMm and LMl project to different parts of the cerebellum (see below, Pakan and Wylie, 2006).

As mentioned above, LM is hypertrophied in hummingbirds (Iwaniuk and Wylie, 2007). In a more recent study, Gaede et al. (2017) recorded from the LM of Anna's hummingbird (Calypte anna) and noted that the neuronal responses to largefield visual stimuli differed in hummingbirds compared with other bird species. First, a directional bias to forward motion was not found, rather all directions were represented equally (Figure 2C). Second, LM neurons in hummingbirds were tightly tuned to stimulus velocity, and mainly to speeds higher than 48°/s (Figure 2E). In this regard, hummingbirds could be regarded as visual “velocity specialist,” forgoing coarse coding common in perceptual systems for a specificity code seen in other sensory specialists, such as neurons tuned for auditory space in owls (Konishi, 2003; Lesica et al., 2010). Insofar as Winterson and Brauth (1985) noted that LMl contained neurons that prefer forward motion, whereas neurons representing all directions of motion, (which prefer faster speeds), were found in LMm, we speculate that the hypertrophy of the LM in hummingbirds may be largely due to an expansion of the LMm.

It is not unreasonable to assume that flight demands acute multisensory integration and sophisticated motor control. The cerebellum is large and well developed in birds (Figure 3A) and as the cerebellum is traditionally implicated in motor control (Ito, 1984) and is clearly a site of multisensory integration (Paulin, 1993; Bower, 1997) it is not surprising that it is thought to be important for flight (Kornhuber, 1974). Husband and Shimizu (2001) noted that the cerebellum is much larger in birds compared to non-avian reptiles (Figures 3A–C). Witmer et al. (2003) examined gross brain morphology from the endocasts of extinct flying reptiles and noted that there was a large cerebellum, in particular, the flocculus (fl) of the VbC (Figure 3D). However, Walsh et al. (2013) urge caution in this regard, because they found no correlation between the flying behavior and the size of the flocculus in extant birds, and large cerebella are also present in several bird-like non-flying dinosaurs. Nonetheless, Walsh et al. acknowledge that the evolution of the large cerebellum may have rendered birds flight ready. Finally, Feenders et al. (2008) used the expression of immediate-early genes as an indicator of neural activity during various behaviors. During flight, activity was evident, especially in folia VI and IXcd. As mentioned, optic flow pathways to the cerebellum in birds are precisely to these two areas, the oculomotor cerebellum (folia VI–VII) and the VbC (folia IXcd and X, see below, Figure 4). Thus, several lines of evidence suggest a key role for the cerebellum in avian flight.

The first optic flow pathway to the cerebellum is to the VbC, folia IXcd and X (Figure 4A). The VbC is divided into the lateral flocculus (fl), which includes the lateral protrusion of the cerebellum known as the auricle (see Figure 3B), and the medial VbC comprised of the ventral uvula and nodulus. Optic flow information reaches the VbC directly, as mossy fibers (MFs) that originate from neurons in both nBOR and LM and terminate in folium IXcd, but not X (see Figures 4A,B for details and the relative position of each nuclei in the brain of the pigeon) (Clarke, 1977; Brecha et al., 1980; Gamlin and Cohen, 1988a). The majority (~75%) of the LM projection is from LMl (Pakan and Wylie, 2006). nBOR and LM also project indirectly to the VbC via the medial column of the inferior olive (mcIO), whose cells ascend to the VbC as climbing fibers (CFs) (Arends and Voogd, 1989; Lau et al., 1998; Wylie, 2001). This indirect projection from the LM arises from fusiform cells that lie in a thin strip along the border between LMm and LMl (Pakan et al., 2006). These two different pathways(CF and MF), also originate from cells with different speed preferences: the CF pathway originates from the slow cells in the nBOR and LM, whereas the MF pathway is fed by both slow and fast cells in nBOR and LM (Winship et al., 2005). This CF pathway to the VbC has been studied in detail in several mammalian species (e.g., Graf et al., 1988), and is critical for retinal image stabilization and the modification of the vestibular ocular reflex by visual-information (Ito, 1998). This is a highly conserved pathway, as the physiology and anatomy of this CF pathway to the flocculus is strikingly similar in birds and mammals (Voogd and Wylie, 2004).

The CFs give rise to the complex spike activity (CSA) of Purkinje cells in mammals, birds and likely other vertebrates (Eccles et al., 1966; Wylie and Frost, 1991). Depicted in Figure 4C, the VbC in birds is organized into several optic flow “zones” (Voogd and Bigaré, 1980), which lie in the sagittal plane and cut across IXcd and X. The CSA in the flocculus responds most strongly to rotational optic flow about either the vertical axis (VA neurons) or an horizontal axis oriented 45° to the midline (HA neurons) (Wylie and Frost, 1993). In pigeon, there are two VA zones interdigitated with two HA zones (Winship and Wylie, 2003). In the uvula/nodulus, the CSA responds most strongly to optic flow resulting from translation (Wylie et al., 1993). As depicted in Figure 4C, there are four response types organized into three sagittal zones. In the most medial zone, CSA responds most stronglyto optic flow resulting from translation backwards along an horizontal axis 45° to the midline such that there is a focus of contraction at 45° contralateral azimuth (contraction neurons). Medial to this is a zone where the CSA responds most strongly to optic flow resulting from either (i) forward translation along an horizontal axis 45° to the midline such that there is a focus of expansion at 45° ipsilateral azimuth, or (ii) upward translation along the vertical axis. Finally, lateral to this is a zone where the CSA responds to the optic flow resulting from downward translation along the vertical axis (Wylie et al., 1998; Graham and Wylie, 2012). Thus, the VbC is well suited to analyze the optic flow resulting from self-translation and self-rotation as birds fly through the world. This is combined with vestibular information derived from separate end organs for analyzing rotation (semicircular canals) and translation (otolith organs) (Pakan et al., 2008).

A second optic flow pathway to the cerebellum of birds is to folia VI-VIII of the posterior lobe, which are collectively known as the “oculomotor cerebellum” (OCb; Voogd and Barmack, 2006). This region of the cerebellum has been implicated in flight insofar as the “strong fliers” as defined by Larsell (1967) tend to have a large posterior lobe, in particular folia VI and VII, and these folia are significantly smaller in flightless birds (Iwaniuk et al., 2007). The LM and nBOR project as MFs to folia VI-VIII (Clarke, 1977; Brecha et al., 1980; Figure 4D). The projection from nBOR is relatively small, but the projection from LM is massive and originates mainly in LMm (~75%) (Pakan and Wylie, 2006). As shown in Figure 4D, folia VI-VII are part of a much more extensive visuo-motor network that incorporates visual information from several indirect sources. These folia also receive heavy MF projections from the medial and lateral pontine nuclei (MP, LP) (Brodal et al., 1950; Freedman et al., 1975; Clarke, 1977; Pakan and Wylie, 2006). The pontine nuclei in turn receive projections from several retino-recipient sources including nBOR (although this is minor; Wylie et al., 1997), the optic tectum, which sends projection mainly to LP (Hunt and Künzle, 1976; Wylie et al., 2009), and LM, which projects mostly to MP (Clarke, 1977; Gamlin and Cohen, 1988a), and Glv (Marín et al., 2001). Glv itself receives input from the optic tectum (Hunt and Künzle, 1976; Wylie et al., 2009; Vega-Zuniga et al., 2014), and in turn projects heavily to LM, mainly to LMm (Vega-Zuniga et al., 2016), providing some interconnectivity among these retinorecipient regions before sending efferents to MP and folia VI-VIII. Because this system receives direct and indirect input from several visual regions, it is well suited for integrating different types of visual information. Whereas the LM is concerned with optic flow, neurons in the tectum respond to small moving stimuli (Frost and Nakayama, 1983; Frost et al., 1990). The exact nature of visual processing in Glv remains elusive, however, electrical stimulation of the Glv elicits discrete, precise orienting head movements (Vega-Zuniga et al., 2011). All of this information is integrated in the pontine nuclei and folia VI-VIII, but for what purpose?

The integration of local motion and optic flow information in primate visual cortex is implicated in “steering” to avoid obstacles during locomotion through cluttered environments (Page and Duffy, 2008; Elder et al., 2009). Similarly, Hellmann et al. (2004) suggested that the tecto-pontine pathway in birds is involved in avoidance behavior. Thus, perhaps this network integrating optic flow and local motion signals is important for obstacle avoidance as birds fly through cluttered environments. For example, during translation, a radial optic flow pattern would result, which would be detected by neurons in LM and nBOR. In addition, self-motion would cause motion parallax of stationary objects at different depths. Such local motion relative to the background, is the ideal stimulus to activate deep tectal cells (Frost and Nakayama, 1983), which project to the pontine nuclei (Hellmann et al., 2004). This combination of local motion and optic flow during self-translation would then be analyzed by the pretecto-ponto-cerebellar system to folia VI-VIII (Figure 4D), allowing a flying bird to then adjust its direction and velocity accordingly.

The visual pathways that convey optic flow information to the cerebellum of mammals are similar to those in birds and have been studied extensively (Simpson, 1984; Voogd and Barmack, 2006). A detailed review of this literature is beyond the scope of this review but a brief summary follows. As in mammals (and other vertebrates), optic flow is analyzed in specialized optic pathways, which begin in two retinorecipient nuclei, the medial and dorsal terminal nuclei of the AOS (homolog to nBOR of birds) and the nucleus of the optic tract (NOT; homolog to LM of birds). These two nuclei are also highly conserved among other vertebrates (Simpson, 1984; McKenna and Wallman, 1985a). Similar to birds, in mammals two regions of the cerebellum ultimately receive visual inputs from these two nuclei, the oculomotor cerebellum (folia VI-VIII) and the VbC (Reviewed in Voogd and Barmack, 2006). However, in contrast to birds, in most mammals optic flow information does not reach the cerebellum directly as mossy fibers from the terminal nuclei and NOT, but rather indirectly through different relay nuclei (see Simpson, 1984; Pakan et al., 2010). As in birds, these nuclei do project to regions of the inferior olive which then projects as climbing fibers to the VbC (Giolli et al., 2006). In the case of the oculomotor cerebellum of mammals, visual projections arise from the pontine nuclei, which like in birds receive projections from the accessory optic system, the NOT, the ventral geniculate, pretectum and the superior colliculus (reviewed in Voogd and Barmack, 2006). The existence of very similar optic flow pathway to the cerebellum of mammals strongly suggest that these are ancestral characters of at least all land vertebrates, and that like the expansion of the cerebellum, they precede the evolution of flight in birds.

Jerison's principle of proper mass states that an increase in the size of a brain nucleus is associated with an increase in the processing power needed to meet behavioral requirement (Jerison, 1973). The classic example is the increased size of the hippocampus in food caching birds (Sherry et al., 1989). Iwaniuk and Wylie (2007) suggested the increase in size of the LM in hummingbirds and other birds is related to an increased need for stabilization during hovering. Although through evolution a single brain nucleus can increase (or decrease) in size, so called “mosaic evolution,” more often large parts of the brain or whole systems evolved together: “concerted evolution” (Striedter, 2005). Gutiérrez-Ibáñez et al. (2014) showed in birds that an increase in size of one visual nucleus is associated with increased in other visual nuclei, but the strength of the correlations varied greatly across pairs of visual nuclei. This is shown in Figure 5, where we show the correlation of the relative sizes of various visual nuclei across 77 species of birds (data are from Gutiérrez-Ibáñez et al. (2014), plus unpublished measurements of pontine nuclei). Note that while the correlation between the LM and nBOR is of a significant magnitude (0.25), the correlation is higher between the LM and PM (0.39) and highest between LM and Glv (0.62). The correlations between LM and PL and LM and the optic tectum are also comparatively low (0.2 and 0.29, respectively). The correlation between PM and Glv is also high (0.54). Thus, the LM tends to increase in size with Glv and the PM, more so than other structure shown in Figures 4A,D. This covariation likely reflects the aforementioned connections among these nuclei and that these regions evolve in a correlated fashion due to increased information processing requirements of one or all nuclei. Based on our findings in hummingbirds, we propose that LM may be the region undergoing direct selection and that the associated changes in Glv and PM are through correlated evolution, as a result of increasing input and output requirements of LM. This preliminary hypothesis could be tested using evolutionary path analysis (von Hardenberg and Gonzalez-Voyer, 2013) to tease apart directionality of correlated size changes.

We have established that LM is enlarged in hummingbirds and, as depicted in Figure 4, that the LM feeds two cerebellar pathways. One is to the VbC (IXcd) that arises mainly from LMl and also from nBOR (Figure 4A). The second is to folia VI-VIII where there is an integration of optic flow signals from LMm and local motion from a tecto-pontine circuit and involving Glv (Figure 4D). We suggest that the hypertrophy of LM in hummingbirds may be due to an expansion LMm given the following facts.

As previously mentioned, Iwaniuk and Wylie (2007) suggested that the hypertrophy of LM in hummingbirds is driven by the increased need to use optic flow to drive the optokinetic response, which facilitates stabilization during hovering. This argument was supported by an expansion of LM in other birds capable of hovering (e.g., kingfishers), although the increase in the size of LM in these species was smaller than in hummingbirds. However, Ibbotson (2017) suggested that the selectivity of hummingbird LM neurons to fast velocities may be related to flying in dense vegetation, as images that are close to the animal will move very quickly across the retina. The five points laid out above suggest that the expansion of LM in hummingbirds could be driven by a differential increase in the size of LMm, where neurons tend to be tuned to higher speeds (Winterson and Brauth, 1985) but also motion parallax, which would occur as they fly through cluttered environments. As Iwaniuk and Wylie (2007) measured LM as whole, there is no evidence that the expansion of LM in hummingbirds is due to a relative increase in the size of LMm vs. LMl. It would be interesting to see if this is the case and if hummingbirds show an increase in the magnitude of the projection from the LMm to the oculomotor cerebellum. In summary, the hypertrophy of LM in hummingbirds may not be related to hovering alone, but could also be related to an increase in the processing demands associated with the pathway to the oculomotor cerebellum as they fly through a cluttered environment while feeding.

In this paper we have reviewed evidence that visual cerebellar pathways are involved in the control of flight. Moreover, there are two pathways involved: a pathway to the VbC involved in the processing of optic flow resulting from self-translation and self-rotation, and a pathway to folia VI-VIII that integrates optic flow and local motion information such as that which occurs during self-motion through cluttered environments. These pathways are fed by different parts of the LM, LMl, and LMm, that have different response properties to optic flow stimuli. Further, we suggest that the hypertrophy of the LM observed in hummingbirds may be driven more so by an hypertrophy of the LMm to support an increased demand of integrating local motion and optic flow signals to process motion parallax in cluttered environment. Hummingbirds spend much of their time feeding, and clearly they will be faced with optic flow and motion parallax as they forage through a patch of flowers. In this respect, we completely concur with Ibbotson (2017) who stated: “It is likely that their [hummingbirds] habit of flying close to flowers in dense vegetation has tuned the hummingbird LM to its specific visual environment.” Hummingbirds may, in fact, be an especially powerful group in which to gain a deeper understanding of the role that optic flow processing plays in flight behavior of birds.

All authors had full access to all of the data in the manuscript and take responsibility for the integrity of the data and the accuracy of the data analysis. Conceived and designed the manuscript: DW. Performed experiments and acquired data: CG-I, AI, and AG. Data analysis: CG-I, AI, and AG. Writing the manuscript: DW, CG-I, AI, and DA. Supervised the study: DW, DA, and AI.