3 Systems and mechanisms for direction selectivity

Since the processing of direction selectivity in the retina is currently a very active research field, there is substantial controversy concerning the relevant entities and activities that contribute to the mechanism, as Anderson points out in his target article. Some accounts focus on local processes within the SAC dendrites themselves (Hausselt et al. 2007), while others draw a broader picture of a multi-component process, where the exact arrangement of cell types and their compartments is vital for direction selectivity (Lee & Zhou 2006). For our purposes here, I would like to use a most recent update on SAC function offered by the group working with Sebastian Seung. The group uses high-resolution electron-microscopy images of brain tissue to reconstruct complete brain networks on a cellular level. Apart from trained reconstruction experts, the project also makes use of so-called “citizen neuroscientists”—volunteers who contribute to the reconstruction process through an online platform that employs gaming features to guide and motivate the community effort (http://www.eyewire.org).

In their study, Seung and colleagues used images from the mouse retina to analyze SAC circuitry. They took a closer look at the exact wiring between bipolar cells (BCs) and SACs (Kim et al. 2014). BCs provide input to SACs, but do not show any directional selectivity by themselves. The main point of the article is to show that different BC subtypes display different patterns of connectivity with SACs. By analyzing branch depth and contact area, they could show that one subtype (BC2) has mainly connections close to the soma, while another subtype (BC3a) has more connections far from the soma in the outer parts of the dendrites. Importantly, the BC subtypes, in turn, have different intrinsic visual response latencies. BC2 seems to lag BC3a by 50ms and more. It can be shown that the differential connectivity patterns and the divergent latencies add up to produce selectivity for a preferred direction of movement going out from the soma on the respective dendrite in accordance with empirical results.

What is important about the paper is not just the main result itself. Any empirical observation may be overruled in the (near) future. So it is not particularly relevant whether these exact cell types and this exact type of wiring is vital for the phenomenon at hand. What I found intriguing in this study, however, was how the relevant mechanism was described and how the data were integrated with a computational model of direction selectivity, reflecting a recent trend in the neurosciences to combine biological and computational perspectives in explanatory accounts. It shows how neuroscientists pick out the relevant parts of a system that contribute to a specific phenomenon in question. The proposed computational model (Fig. 1a; Kim et al. 2014) maps the biological entities onto specific parts of the computational circuit. The output element at the lower part of the figure is the SAC. The input stems from BC2 (left) and BC3a (right); their respective response properties are captured as delay values and sustained vs. transient response types. The circuit combines elements of classical models of direction selectivity, the Reichardt (Fig. 1b) and the Barlow-Levick detectors (Fig. 1c). Clearly, the direction selectivity cannot be attributed to any one of the system components in isolation. Mechanistic accounts and the corresponding computational models both point to the whole complex of entities as the relevant system that achieves directional selectivity.

Figure 1: Computational models of direction selectivity (a) The selectivity of SACs described in Kim et al. (2014) can be modeled with a computational framework using a combination of sustained and transient response properties as well as excitatory and inhibitory connections. The displayed wiring would lead to direction selectivity for rightwards motion. The proposed model can be considered to combine previous classical models of direction selectivity, the Reichardt detector (b) and the Barlow-Levick model (c). Green dots indicate excitatory and red dots inhibitory synapses. ‘-τ’ indicates a temporal lead and ‘+τ’ a temporal lag. Reprinted by permission from Macmillan Publishers Ltd: Nature (Kim et al. 2014), copyright (2014).

In its computational abstraction, the model can be thought of as a canonical system of directional selectivity. Similar models have also been applied to different hierarchical levels of neural processing and different species. For example, mechanisms of directional selectivity have been studied for a long time in the fly visual system. With very different neural elements and wiring, a system of interconnected neurons achieves directional selectivity with response properties closely resembling the Reichardt-type of motion detector (Borst & Euler 2011). Again, only the combination of elements from different processing stages succeeds in delivering direction selectivity as a system. On a cortical level, direction selectivity has been first described for complex cells of the primary visual cortex (V1) in the seminal work of David Hubel & Torsten Wiesel (1962). Without offering a quantitative computational model, they nevertheless suggest a hypothetical connectivity pattern between different cell types that might underlie the observed responses to moving patterns in complex cells (Hubel & Wiesel 1962, Fig. 20). The model shares features with other motion detectors; a mapping between components is possible.

When it comes to motion selectivity in the brain, one of the most intensively studied cortical areas is the middle temporal (MT) region. The region was first described in the macaque (Dubner & Zeki 1971; Zeki 1974) and owl monkeys (Allman & Kaas 1971). The human homolog, the human MT complex (hMT+; Tootell et al. 1995; Zeki et al. 1991), turned out to be a collection of areas with related response properties (Amano et al. 2009; Kolster et al. 2010). Again, to understand the direction selectivity of MT, it is necessary to consider the cooperation of cells in MT and the input processing stages, mainly from V1. This cooperation and the need for an integrated perspective is emphasized in empirical studies (Saproo & Serences 2014) as well as computational models of MT functioning (Rust et al. 2006). Only the V1-MT system as a whole is understood to deliver motion selectivity as output of the MT stage.

But in terms of the role of MT in motion processing, a case could be made in support of Anderson’s suggested distinction between a system that exhibits a certain selectivity and the mechanism that produces this selectivity. The apparent locality and modularity of motion processing in MT is based on very selective deficits in patients with lesions in and around MT (Zeki 1991; Zihl et al. 1983). And stimulation of MT with transcranial magnetic stimulation (TMS) in healthy participants leads to selective deficits in motion perception (Beckers & Hömberg 1992; Beckers & Zeki 1995; Hotson et al. 1994; Sack et al. 2006). In a recent study, patients undergoing brain surgery near MT could be investigated with electrical stimulation (Becker et al. 2013). Only stimulation of MT and a related area nearby, MST, led to an inability to perform a simple motion-detection task, a rather specific result concerning the relevance. Results of that kind drive the intuition that the system that is responsible for motion perception, independent of any cortical areas that might mechanistically contribute to the processing chain leading up to MT (like V1), are localized in MT.

Lesion and other interference studies (e.g., with TMS) are suggestive, but there are also well-known difficulties with interpreting the results. Lesions mostly affect larger parts of the brain and are rarely limited to a single cortical site. As such it is often hard to identify the actual parts of the complex brain networks that are affected. The advantage of stimulation techniques is that the interference is temporary and can be precisely targeted on a specific location. But, given the rich connectivity structure of neural networks, stimulation effects can be seen even in remote target sites (Bestmann et al. 2004; Sack et al. 2007). In addition, TMS studies have shown that activity of MT might not even be sufficient for conscious motion perception without the involvement of V1 (Pascual-Leone & Walsh 2001; Silvanto et al. 2005). There are also further empirical as well as philosophical reasons for rejecting the claim that motion perception can be attributed to MT in a stringent fashion (Madary 2013), which I won’t discuss here.[1]

So while at first glance MT is a very strong candidate for straightforward and very local attribution of function, it seems again that the relevant system is more appropriately described on a network level. The tendency to see system parts as vital for a function may also stem from the limitations of our employed methods. Lesion cases and interference techniques are commonly interpreted as being informative about the relevant gray-matter structures that are affected by the lesion or stimulation. But there is evidence that interference with white-matter connections between network parts can be even more incapacitating than gray-matter damage. It has long been known that frontoparietal areas are implicated in a deficit of visuospatial attention called neglect. But very recently Thiebaut de Schotten et al. (2005, 2011) revealed that the properties of fiber connections between frontal and parietal sites are most predictive of visuospatial processing capacities, and that their electrical stimulation leads to severe deficits. Transferring this insight to the case of MT, we simply have most direct access to the cortical gray-matter centers involved in motion processing, and since they are vital components of the system, this also leads to corresponding deficits when they are affected or stimulated. But this might conceal the fact that motion selectivity is a product of a wider network that crucially depends on integrated processing for proper functioning.

In sum, I think that close inspection of how direction selectivity is investigated and treated in neuroscientific research is in disagreement with Anderson’s arguments (1) and (3). Although it is true that investigators sometimes refer loosely to local elements as displaying a certain characteristic, the corresponding detailed and extended accounts of direction selectivity give credit to the distributed nature of the relevant systems that figure in explanations. Even considering the case of conscious motion perception, it is unclear whether the presumed locality of motion representation stands up to stringent tests. Rather, it seems to be a case of localized interference with a distributed system where damage to vital hubs leads to fundamental deficits.