Target Selection by the Frontal Cortex during Coordinated Saccadic and Smooth Pursuit Eye Movements

The primate retina affords wide-field detection of visual pattern and motion, as well as narrow-field scrutiny of visual details via its central zone, the fovea. A typical oculomotor episode involves detection of visual stimuli by receptors in the retinal periphery, cortical selection of one of the detected stimuli as a goal for scrutiny, and genesis of eye movements that enable foveation of the goal. Two eye movement systems cooperate to achieve foveation if the goal stimulus is moving: the saccadic (SAC) and the smooth pursuit eye movement (SPEM). The SAC eye movement system generates ballistic, high-velocity, open-loop ocular rotations that quickly cancel the difference between the initial angle of gaze and the angle needed to foveate the goal. The SPEM system generates continuous, moderate-velocity, closed-loop ocular rotations that prolong foveation of mobile goal stimuli by trying to match gaze velocity to stimulus velocity. For sufficiently fast stimuli, prolonged foveation requires re-engagement of the SAC system to generate “catch-up” saccades.

It is critical that the two eye movements systems coordinate so that they (1) choose the same target as the goal for future movement, and (2) maximize visibility of that target. A prior report (Grossberg, Srihasam, & Bullock, submitted) introduced a new model of SAC–SPEM interactions, and reported data-simulation comparisons that revealed how the model achieved the second of these requirements. This report presents additional data-simulation comparisons which illustrate how circuit interactions ensure that both systems generate movements directed to the same target.

It might be thought that unitary targeting is not a significant problem. However, one system can operate without the other, a typical scene contains an assortment of salient stimuli with some stationary and some moving, and the motion-sensitive anatomical pathways that generate SPEM are largely separate from, and evolved after, those that generate SAC (Krauzlis, 2005; Krauzlis, Liston, & Carello, 2004). What interactions choose a unitary goal among several competing stimuli? Is goal selection information transferred from one system to the other, and if so, how?

Although SAC and SPEM systems have often been assumed to have a single, shared, target selection mechanism, recent data from monkeys suggest that each system has a selection mechanism that can operate independently of the other. In these experiments, two different stimuli appear and begin to move in different directions. Location, shape, direction of motion, or color help the monkey distinguish a task-relevant target from an irrelevant distractor (Garbutt & Lisberger, 2006; Adler, Bala, & Krauzlis, 2002; Gardner & Lisberger, 2001; Krauzlis, Zivotofsky, & Miles, 1999; Ferrera & Lisberger, 1995). Because the latency for SPEM initiation is shorter than for SAC initiation, it is possible to behaviorally assess whether the SPEM system always reflects the same choice as the slower-starting SAC system. Indeed, on a substantial fraction of trials, the initial SPEM tracks the distractor. However, immediately after a saccade to the correct target, the continuing SPEM tracks that target rather than the distractor. Such results suggest that each of the SAC and SPEM systems makes its own initial choice of goal, but that the SAC choice overrides the SPEM system's choice (Adler et al., 2002; Gardner & Lisberger, 2001, 2002; Horwitz & Newsome, 2001; Ferrera & Lisberger, 1995).

Before the model of Grossberg et al. (submitted), no model had addressed the functional anatomy of SAC–SPEM interactions for target selection and coordination. Although a lot is known about the mechanisms involved in saccadic target selection (Thomas & Pare, 2007; Schiller & Tehovnik, 2003, 2005; Arai, McPeek, & Keller, 2004; Krauzlis et al., 2004; Basso & Wurtz, 2002; McPeek & Keller, 2002; Thompson, Hanes, Bichot, & Schall, 1996; Schall, Hanes, Thompson, & King, 1995), less is known about how smooth-pursuit target selection occurs (Case & Ferrera, 2007; Garbutt & Lisberger, 2006; Krauzlis, 2005), or about how these two eye movement systems interact to select the same target among many distractors (Adler et al., 2002; Gardner & Lisberger, 2001, 2002; Krauzlis et al., 1999). Our new model proposes a mechanistic explanation for, and quantitatively simulates, the key generalizations that have emerged from systematic empirical studies of target tracking eye movements concerning how the smooth pursuit system selects its target, and how the two eye movement systems interact to select the same target. After reviewing the SAC–SPEM model's circuitry and operation, the present article explains how model circuits, notably those dependent on the frontal cortex, can explain and simulate challenging behavioral and electrophysiological data from experiments on target selection.

The model (Figure 1) consists of two parallel yet interacting processing streams for the control of SPEM and SAC movements. Neuroanatomy-based models of SPEM and SAC eye movements were used as starting points for this work: the SACCART model of learning and performance using a multimodal movement map in the superior colliculus (SC; Gancarz & Grossberg, 1999; Grossberg, Roberts, Aguilar, & Bullock, 1997), the TELOS model of learned cue-guided voluntary selection of saccade plans (Brown, Bullock, & Grossberg, 2004), and a model of motion perception and SPEM command genesis in the middle superior temporal area (MST; Pack, Grossberg, & Mingolla, 2001). A complete mathematical specification of the model, and details regarding simulations, are provided in the Supplementary Material section.

Figure 1A gives an overview of the structure of the model SPEM circuit. The model's smooth pursuit stream contains visual area middle temporal (MT)-like cell types, which are distinguished by selectivity for different combinations of the direction and speed of visual stimuli that fall within their retinotopic receptive fields (Albright, 1984; Maunsell & van Essen, 1983a). The 800 model MT cells provide inputs to the model's MST cells, which pool their MT inputs in a way that makes them direction-selective and speed-sensitive, but not speed-selective. Because these MST cells also receive inputs corresponding to current eye velocity, they can compute an internal estimate of predicted target velocity that remains accurate even as eye velocity grows to match target velocity, and thus, gradually cancels the target-related retinal image motion that drives MT cells.

The predictive estimate of target velocity computed by model MST cells provides a basis for the model's frontal cortical representation of desired pursuit velocity. In particular, the frontal pursuit area (FPA), at the rostral bank of the arcuate sulcus, receives inputs from both MST and MT (Tian & Lynch, 1996a, 1996b; Huerta, Krubitzer, & Kaas, 1987). Model and real FPA cells have high direction selectivity and speed sensitivity, but almost no speed selectivity (Tanaka & Lisberger, 2002b; Gottlieb, Bruce, & MacAvoy, 1993).

The model FPA cells project (Giolli et al., 2001; Brodal, 1980) to the model nucleus reticularis tegmenti pontis (NRTP), which includes two types of cells: acceleration cells and velocity cells (Ono, Das, Economides, & Mustari, 2005; Ono, Das, & Mustari, 2004; Suzuki, Yamada, Hoedema, & Yee, 1999; Yamada, Suzuki, & Yee, 1996). Model NRTP velocity cells integrate the output of NRTP acceleration cells. The latter compute the difference between an excitatory target velocity command from FPA and an inhibitory eye-velocity signal from the vestibular nuclei. This difference estimates the eye acceleration that is needed to match target velocity. These two classes of cells allow the NRTP to play a key role in SPEM initiation.

One possible source of this inhibitory eye velocity signal to NRTP might be from the ventrolateral thalamus. Neurons in the ventrolateral thalamus discharged before or during initiation of pursuit, and the firing rate was proportional to the speed of target motion in a preferred direction. When the tracking target was extinguished briefly during maintenance of pursuit, these neurons continued firing, indicating that they carried extraretinal eye movement signals (Tanaka, 2005).

Parallel to the FPA–NRTP pathway, a second pathway exists for the transmission of SPEM-related information from the cortex to the cerebellum (CBM) via the pons. This pathway, apart from being implicated in maintenance of SPEM (Mustari, Fuchs, & Wallman, 1988; Suzuki & Keller, 1984), also helps to program saccades made to moving objects. Model MT cells project to the dorsal lateral pontine nucleus (DLPN) cells of the brain stem, which then project to the CBM (Figure 1A). In the model, the DLPN cells have similar speed and directional selectivities as MT cells, but they lack their retinotopic specificity.

In the model SAC system, retinotopically organized visual signals are processed to produce saccadic target choices in the model lateral intraparietal area (LIP) and the frontal eye fields (FEF). FEF outputs serve as inputs to corresponding retinotopic loci in the burst and buildup cell layers (Sommer & Wurtz, 2000, 2004; Munoz & Wurtz, 1995a, 1995b) of the motor error map of the model's SC. There is also communication between these SC layers, such that activated loci in the burst cell layer excite corresponding cells in the buildup cell layer (cf. Grossberg et al., 1997).

Outputs from the SC reach the CBM and the saccade generator circuit (Figure 1B) in the paramedian pontine reticular formation (PPRF), which contains populations of SAC- and SPEM-related cells, some of which provide direct input to the oculomotor neurons that innervate eye muscles. Model saccadic control signals from cerebellar and SC stages converge at model long-lead burst neurons (LLBN). LLBN activity encodes gaze-position error. LLBN cells excite corresponding excitatory burst neurons (EBN), which project to the tonic neurons (TN). TNs integrate inputs from EBNs and excite the model oculomotor neurons. The EBNs also excite inhibitory burst neurons (IBN), which in turn inhibit the LLBNs, thereby completing an internal negative feedback loop that controls ballistic saccades. Except during saccades, the EBNs receive strong inhibition from model omni-pause neurons (OPNs), so-called because they pause deeply to disinhibit saccades of all directions. For reviews of these anatomical connections and neurophysiological properties, see Büttner & Büttner-Ennever, 1989, 2005; Moschovakis & Highstein, 1994; Fuchs, Kaneko, & Scudder, 1985; for reviews of computational models that incorporate such features, see Girard & Berthoz, 2005).

OPNs are located in the nucleus raphe interpositus. The pursuit neurons (PNs) found in the vestibular nuclei are modeled as receiving input from the CBM and projecting directly to the TNs, which are thus shared by SAC and SPEM systems. The PNs are weakly inhibited by, and themselves inhibit, the OPNs, also shared by both systems. About 50% of the OPNs show reduced activity (a 34% drop) during smooth pursuit (Missal & Keller, 2002), whereas most OPNs pause more deeply during saccades (Munoz, Dorris, Pare, & Everling, 2000; Everling, Pare, Dorris, & Munoz, 1998). Thus, the spontaneously active and inhibitory OPNs normally oppose both saccades and SPEM. Shallow pausing by OPNs can release SPEM but not saccades, whose release requires deeper pauses.

Learning is needed to keep SAC and SPEM metrics accurate as eye muscles and other system parameters change. Inactivation or lesion of the CBM causes deficits in the ability to adapt both SAC and SPEM (Takagi, Zee, & Tamargo, 1998). Each model cell in the retinotopic SC, the speed-sensitive DLPN, and the direction-sensitive NRTP sends signals to the CBM (Thier & Ilg, 2005). These signals are modified by adaptive weights learned within the CBM. The adaptive saccade-related cerebellar outputs reach the model PPRF region of the brain stem (Figure 1A), the location of the saccade generator (Figure 1B). Similarly, the weighed pursuit-related cerebellar outputs reach the PNs in the vestibular nuclei of the brain stem.

Behavioral data (simulated in Grossberg et al., submitted) suggest the existence of an intelligent mechanism to control saccade initiations during SPEM, notably to inhibit saccades when targets are already foveal or parafoveal. Pursuit-related neural activity is reliably observed in the SC: Rostral parts of SC (rSC) contain cells that respond to both SPEM and SAC eye movements (Krauzlis, Basso, & Wurtz, 2000). As schematized in Figure 2, area MT sends strong excitatory projections to the rSC (Collins, Lyon, & Kaas, 2005; Maioli, Domeniconi, Squatrito, & Riva Sanseverino, 1992; Davidson & Bender, 1991; Wall, Symonds, & Kaas, 1982; Spatz & Tigges, 1973), which in turn provides the main excitatory input to the OPNs (Everling et al., 1998; Gandhi & Keller, 1997; Pare & Guitton, 1994; Munoz & Wurtz, 1993a, 1993b). Our hypothesis that rSC projects to the OPN is consistent with data from Büttner-Ennever, Horn, Henn, and Cohen (1999), which state that “there are multiple projections directly onto OPNs from the rostral SC but not from the caudal SC associated with large gaze shifts….” Also, the place where FEF fibers terminate appears identical to the region where SC fibers terminate (Büttner & Büttner-Ennever, 1989). Stanton, Goldberg, and Bruce (1988) found direct projections from the fundus of the arcuate sulcus (FEF) to a region in nucleus ralphe (OPNs) identical to where SC fibers terminate. Thus, we hypothesize that OPNs might be getting foveal input from both the FEF and the SC. In the model (Figures 1 and 2), foveal and parafoveal cells in area MT, which are active when pursued targets are on or near the fovea, inhibit saccades via an excitatory pathway from MT to rSC to the OPNs.

Selecting a saccadic target present among competing distracters, as in a visual search paradigm (Treisman & Gormican, 1988), requires dissociation of target stimuli from others. Visually responsive neurons in the FEF show this discrimination. Late phase activity of these neurons was accentuated for targets and attenuated for distractors that excited their motor receptive fields (Schall & Thompson, 1999; Thompson, Bichot, & Schall, 1997; Thompson et al., 1996; Schall, Hanes, et al., 1995). Stimulation of an LIP, FEF, or SC neuron increases the probability of choosing the stimulus within its motor field as the target (Carello & Krauzlis, 2004; Schiller & Tehovnik, 2001, 2003), and reversible inactivation of the FEF, LIP, and SC produces significant saccadic target selection deficits (McPeek & Keller, 2004; Schiller & Tehovnik, 2003; Wardak, Olivier, & Duhamel, 2002).

In the model, interactions in the parieto-frontal circuit (LIP–FEF–LIP) are critical for selecting targets among competing distractors. Figure 3 schematizes interactions in the SAC pathway that lead to target selection. FEF planning neurons (similar to the FEF visual motor cells in Schall, Hanes, et al., 1995), which receive input from lower-level visual areas (not shown), send outputs to the LIP, FEF output neurons (similar to the FEF motor cells in Schall, Hanes, et al., 1995), SC buildup neurons, and the striatum of the basal ganglia (BG; Schall, Morel, King, & Bullier, 1995). In amphibians and all land vertebrates, the BG interact with the optic tectum (OT), or its homologue, the SC, to control orienting action (Marin, Smeets, & Gonzalez, 1998; Butler & Hodos, 1996). In mammals, the BG interact with frontal cortical areas to control orienting, cognitive, and manipulative behaviors (Strick, Dum, & Picard, 1995; Passingham, 1993; Hikosaka & Wurtz, 1983). Lesions of the BG cause disorders such as Parkinsonian akinesia, Huntington's chorea, and ballism (Albin, Young, & Penney, 1989), suggesting a strong link between the BG and areas involved in movement control. These interactions provide a natural basis for differentiating plan activation and plan execution. Brown et al. (2004) developed a detailed neural model of how the FEF and the BG interact with movement-controlling areas such as the SC.

The current SAC–SPEM model approximates BG-mediated decision-making by sending a GO signal excitation to the FEF if the sum of the FEF, LIP, and SC signals associated with a particular saccadic vector is large enough to exceed a threshold. In vivo, this signal results from striatal disinhibition of the thalamus, and can be regarded as opening a normally closed gate to release the planned movement (Lo & Wang, 2006; Brown et al., 2004; Wurtz & Hikosaka, 1986). Once the gate is open, the FEF cell having maximal activity is selected as the target, and it suppresses other target representations via mutual inhibition.

We predict that the smooth eye movement part of the FEF, known as the FPA, is involved in decision-making for SPEM and for coordinated tracking by SPEM and SAC. The FPA, located near the rostral bank of the arcuate sulcus, is strongly innervated by the MST (Tian & Lynch, 1996a, 1996b; Huerta et al., 1987). FPA cells have high direction-tuning and almost no speed selectivity (Lynch & Tian, 2005; Tian & Lynch, 1996a, 1996b, 1997; Gottlieb et al., 1993). Lesions in the FPA cause SPEM deficits, including reduction of pursuit gain and directional deficits (Shi, Friedman, & Bruce, 1998; Keating, Pierre, & Chopra, 1996). Electrical stimulation of the FPA causes SPEM, and, if applied during natural SPEM, such stimulation increases SPEM gain if the natural pursuit direction and the stimulated direction are the same (Carey & Lisberger, 2004; Tanaka & Lisberger, 2002a, 2002b, 2002c; Gottlieb et al., 1993).

In the model, an SPEM target choice can be made independently by the FPA under the influence of visual motion inputs from the MT and the MST. Directions of all the moving objects in the environment form retinotopically organized peaks of activity of neuronal populations in the MT. We hypothesize that the FPA, which lies downstream from the MT and the MST, converts this sensory signal to a motor signal for movement. This assumption is in agreement with data showing multiple peaks in the MT, before selection occurs downstream (Treue, Hol, & Rauber, 2000). This fact is reflected in model dynamics (compare Figure 7E and F below). Related data about selection of saccadic movement commands beyond MT can be found in Roitman and Shadlen (2002) and Shadlen and Newsome (2001), and are modeled in Grossberg and Pilly (2008).

A GO signal that represents the excitatory thalamo-cortical output of the BG–thalamus system is released when the FPA signal to the BG exceeds a threshold for movement initiation. This assumption accords with recent data (Cui, Yan, & Lynch, 2003) indicating that the FPA connects with the BG–thalamus system in a way that parallels FEF connections with that system (see Brown et al., 2004).

Although both SAC and SPEM systems can reach decisions, the model also provides for coordination. Krauzlis et al. (1999) found that once a target has been selected by the saccadic system, and a saccade is initiated, it is rare for the eye to move along the direction of any contending stimulus after the saccade. Figure 1A shows that there is a path in the model (and in vivo) from the FEF to the LIP to the MT to the MST to the FPA. This model pathway mediates transfer of the SAC system's target choice to the SPEM system. This transfer enables the SAC system to override the SPEM system. Although the FEF and the FPA are nearby in the frontal cortex, the model links them via other areas, and not directly, in accord with both computational and anatomical constraints.

The LIP is key in that linkage. First, it receives collaterals from both saccadic and smooth pursuit pathways (Tian & Lynch, 1996a, 1996b; Schall, Morel, et al., 1995; Maunsell & Van Essen, 1983b). Second, its activity strongly reflects a monkey's decision about a subsequent eye movement, and choice-related up-modulation of direction-tuned cells are seen regardless of whether the sensory inputs driving this decision are stationary or moving stimuli (Schiller & Tehovnik, 2005; Gold & Shadlen, 2000, 2001; Shadlen & Newsome, 2001). Grossberg and Pilly (2008) describe a model for how motion stimuli influence LIP decision-making. Third, activation of LIP affects target choice without affecting the saccade metrics such as saccade duration or velocity (Thomas & Pare, 2007; Schiller & Tehovnik, 2005; Wardak et al., 2002; Schiller & Chou, 1998).

The Supplementary Material section provides simulation details, including the system of differential equations that, together with external inputs, govern the model circuit's dynamics.

Consider the case of selecting a saccadic target among two stimuli (see Figure 4A). The initial phase of activity (50–100 msec) is similar for both target and distractor (see Figure 4B, C, and D), consistent with observations that visual neurons of the FEF show transient activity as early as 50 msec after stimulus onset (Schall, Hanes, et al., 1995). But the second phase of activity is selective for the target. FEF planning neurons representing the target increase, and other neurons representing distractors decrease, in their activity (Figure 4B). Because the two stimuli are identical, there is an equal probability for either to be chosen as target. The choice is the result of a competitive race between the cells coding the two stimuli. Attentional bias, in the form of oddball size, shape, or color stimulus, can help break the symmetry and determine the winner. In the present case, no such bias was present, so noise was added at the FEF stage. After an interval of slow separation between the competing representations, which begins in FEF cells at the time marked by arrows in Figures 4B and D, the BG–thalamus stage reaches threshold. Its output (GO signal) to FEF initiates the rapid separation that is visible first in FEF planning and output cells (Figure 4B and C), and after a brief delay, in LIP cells (Figure 4C, arrow). Thus, the model predicts that the choice is made by the frontal circuit and rapidly relayed to the LIP. Cortical lateral inhibition mediates the model's rapid suppression of the unchosen representation.

Figure 5 compares the simulated activity of model FEF planning and output neurons with FEF electrophysiological recordings (Schall, Hanes, et al., 1995) from a visual search experiment in which targets were clearly discriminable from distractors. In the simulation, this difference was mimicked by making one stimulus 1.5 times as strong as another one. The stronger stimulus was always chosen, and the activation profiles in the model are qualitatively similar to the recordings.

Often, one of several moving stimuli is chosen for tracking by SPEM, in the absence of any saccades. If two moving stimuli are nearby the initial gaze position, then the direction of this SPEM often begins as the vector average of the two stimulus motion directions, then quickly evolves into pursuit along the direction of one stimulus. The second phase reflects a decision, which is also signaled by an increase in SPEM gain: The ratio of eye velocity to target velocity (SPEM gain) is much lower during the initial vector averaging phase than after the system has committed to pursuit of one target. Figure 6 highlights the idea that the phase transition occurs once a decision process in the BG–thalamus exceeds the threshold for generating a GO signal that increases the output from FPA averaging cells to subcortical SPEM generators.

Figure 7 presents simulation results that illustrate the model's ability to generate a similar transition between vector averaging and choice of one target for high-gain SPEM, and also shows the consequence of the decision for the activity of model FPA output cells. Unlike saccadic vector averaging, which holds for an entire saccade, and occurs only when there is small angular separation between the potential targets and when SAC latency is very short (Arai et al., 2004; Ottes, Van Gisbergen, & Eggermont, 1984), pursuit averaging is more commonly observed (Gardner & Lisberger, 2001, 2002; Tanaka & Lisberger, 2002c; Recanzone & Wurtz, 1999). In the model, the difference between the two systems occurs because the output cells in the FPA can become partly activated before a decision is reached, unlike output cells of the FEF. Such a lower threshold for output from the FPA than the FEF makes behavioral sense because averaging low-gain SPEM incurs no visual cost, whereas a premature saccade would impose costs in the form of an interval of reduced vision and defoveation of the target.

Figure 6 shows that two further classes of model FPA cells, which correspond to what Tanaka and Lisberger (2002c) called “summation” and “WTA” (winner-take-all; see Grossberg, 1973, 1982, for how to design networks with a WTA property) cells, receive inputs from model areas MST_d and MST_v, respectively, and excite FPA output (averaging) cells. Figure 8A shows data on all three types of FPA cells, and Figure 8B shows how the model simulates the qualitative differences between these cell types under the three conditions tested. These patterns illustrate that both summation cells and WTA cells come close to their maximum activities levels during the vector averaging phase (i.e., before a choice has been made), whereas the output (averaging) cells show much lower than maximal activity until the disappearance of one target results in choosing the one that remains. This property suggests that the averaging cells are output cells that reflect the BG–thalamus decision process.

Note that summation cells show a decline in activation during SPEM maintenance (in the latter part of the trial), whereas WTA cells show sustained activity during this interval. This difference is explicable in the model because summation cells receive input from area MST_d cells, which become much less active as successful SPEM cancels target motion signals in the retinal frame (since there are no background stimuli), and because WTA cells receive inputs from MST_v cells, which remain active even during successful pursuit because they receive a corollary discharge of eye velocity (Pack et al., 2001).

If the FPA is involved in deciding which stimulus to follow, then stimulation of the FPA should bias the result during target selection. Because microstimulation in the FPA causes eye movements along the particular direction represented at the stimulation site, any stimulus moving along this direction should be favored and selected over competing stimuli. Data of Tanaka and Lisberger (2), Figure 6) confirm this idea. The simulation results plotted in Figure 8C and D show that this is also true of the model. Initially, there are two stimuli moving in opposite directions (left and right directions in our simulations) and toward the fixation point. Selective stimulation of left-tuned model FPA output cells causes leftward (plotted as downward) SPEM. At 200 msec after the stimulus turns on, one of the stimuli disappears (in Figure 8C, the left-moving stimulus disappears; in Figure 8D, the right-moving stimulus disappears). At the same time, the stimulation to the FPA ceases. Figure 8C exemplifies a case in which the stimulated direction does not match that of the target that remains after stimulation offset. Following stimulation offset, the eye stops moving leftward, makes a rightward catch-up saccade, and then pursues the right-moving target. Figure 8D illustrates what happens when the electrically stimulated direction is the same as the direction of the remaining target. Although the SPEM latency is the same for stimulation and nonstimulation cases, the eye reaches target velocity faster when the FPA is stimulated.

The latency of SPEM initiation depends on the distribution of directions of moving stimuli. Adding a second moving stimulus will increase or decrease the latency of SPEM initiation if the second stimulus moves, respectively, in the same or the opposite direction as the chosen target (Ferrera & Lisberger, 1995). Note that in the case of oppositely moving stimuli, the vector average will be no movement. In such cases, no SPEM will be observed until the system truly makes a choice. In the model, stimuli moving in different directions activate cells in the MT and the MST that are tuned to these directions. Such direction-sensitive cells project to the CBM and later output stages via two pathways (Figure 1A): one via the DLPN and another via the FPA and the NRTP. Although both branches can contribute to short-latency SPEM via vector averaging, neither will do so when the average of two opposing motions is zero. But if two potential targets are moving in the same direction, MT, MST, and FPA summation cells coding this direction will often be more active, notably if there are two motion inputs within their receptive fields. The summation cells of the FPA almost double their firing rate (Figure 8A, middle panel) if they contain two targets in their receptive field. This leads to faster activation of FPA output cells. Faster activation of FPA output cells leads to faster CBM cell activation, and thus, to faster pursuit initiation. Figure 9 illustrates these effects and shows model simulations (right) next to behavioral data (Ferrera & Lisberger, 1995).

Behavioral and neurophysiological data indicate that target selection can occur independently in the SAC and SPEM subsystems, as simulated above. Independent selection raises a coordination problem: Which choice dominates in a tracking episode, and how is the coordination achieved? We hypothesize that SAC choice overrides SPEM choice. Data from behavioral (Adler et al., 2002) and microstimulation (Gardner & Lisberger, 2001, 2002) experiments support this claim. Figure 10 compares model simulations (right) of the Gardner and Lisberger (2002) paradigm with their published data (left). In this double-stimulus paradigm, where natural saccades to either stimulus are equiprobable, an FEF site with a known directional preference and movement receptive field is stimulated after pursuit initiation, just as one stimulus enters the movement field, but before any natural saccades would occur. The stimulation evokes a saccade to the stimulus, even if the presaccadic SPEM was tracking the other stimulus. Moreover, SPEM velocity after the evoked saccade almost always matches the velocity of the saccadic target, not the velocity of the stimulus being tracked before the evoked saccade. Thus, SAC choice overrides an existing SPEM choice. In the model, these properties emerge because focal FEF stimulation activates a corresponding retinotopic zone in the LIP which, in turn, activates a retinotopic zone in the MT and the MST, within which inputs from the motion of the favored target will activate the correct direction-tuned cells that project to later stages of the SPEM system (Figures 1 and 6).

Figure 11 shows the predicted activation dynamics for two competing cells at each of five stages of the model during an episode in which an SAC choice overrides an SPEM choice. The series of panels, from top to bottom, corresponds to the path of information transfer, across involved stages, for communicating the choice from the SAC to the SPEM system. Thus, Figure 11A plots FEF planning neurons; Figure 11B, LIP saccadic neurons; Figure 11C, MST_v direction cells; and Figure 11D, FPA output averaging cells. Solid lines represent the traces of cells whose preferred directions match the stimulus that is eventually chosen as target by the SAC system. In Figure 11D, the FPA stage of the SPEM system initially chooses the other stimulus (dashed trace; see first arrow in Figure 11D). Note that the time at which the FPA flips its choice (second arrow in Figure 11D) is mere milliseconds after the FEF makes its choice. This allows the SPEM system to establish a good representation of the new target's velocity before the saccade is actually generated. This enables the postsaccadic SPEM to exhibit a velocity that is already matched to the target velocity.

Primates use a combination of SAC and SPEM to track a selectively attended moving object. Recent data on monkeys performing a target selection task indicate that each system has its own selection mechanism, but that in the event of a conflict, the SAC system's choice overrides the choice made by the SPEM system (Adler et al., 2002; Gardner & Lisberger, 2001, 2002; Horwitz & Newsome, 2001; Ferrera & Lisberger, 1995). This article and its companion (Grossberg et al., submitted) describe a neural circuit model that is capable of simulating these and a large number of additional data concerning how the SAC and SPEM systems interact to select and track targets among distractors in a coordinated way.

In our model, saccadic target selection emerges from interactions of frontal, parietal, and collicular regions with the BG. Visually responsive cells in the FEF, SC, and LIP show target/distractor discrimination at about the same time after target onset: FEF, 130–150 msec (Thompson et al., 1996; Schall, Hanes, et al., 1995); SC, 110–130 msec (McPeek & Keller, 2002); and LIP, 132 ± 2.3 msec (Thomas & Pare, 2007). However, when compared with respect to saccadic initiation time, t = 0, larger population of FEF and SC cells show earlier activity: FEF, 78% at t = −53 msec; SC, 98% at t = −45 msec; LIP, 60% at t = −26 msec. This accords with our claim that the FEF and the SC plan and execute a saccade and this information is passed to the LIP to coordinate and plan further eye movements.

The anatomical loci and processes that are most critical for SPEM target selection are still unsettled. Our model is based on the hypothesis of a partial parallelism between the SAC system and the SPEM system. Notably, SPEM target selection emerges as a result of interactions between the BG and a frontal area, the FPA. Stimulation of the FPA, or any other region downstream from it in the SPEM circuit, notably NRTP or a pursuit zone of the CBM, causes SPEM within a very short latency: FPA, 25–35 msec (Tanaka & Lisberger, 2002b; Gottlieb et al., 1993); NRTP, 20 msec (Yamada et al., 1996); CBM, 10 msec (Belknap & Noda, 1987). This is also true in the model. However, stimulating upstream from the FEF, at the MST or MT, modulates existing pursuit but never evokes SPEM from a voluntarily maintained fixation (Komatsu & Wurtz, 1989). The same would be true in the model if it were augmented to include a STOP signal channel to complement the GO signal channel. Such a STOP signal channel could suppress voluntary fixation responses to MT and MST stimulation in a manner analogous to that shown for the SAC system in Brown et al. (2004). These considerations reinforce the hypothesis of partial SAC–SPEM parallelism, and the more specific hypothesis that the voluntary gating associated with SPEM target selection occurs in the FPA, which connects with the BG–thalamus system in a way that is analogous to the FEF.

Before noting the limits of the parallelism in our model, it is useful to consider and evaluate an alternative view. There is growing evidence for rSC involvement in pursuit. Some nonfixation neurons in the rSC show target/distractor discrimination. Stimulation of rSC inhibited pursuit in the ipsilateral visual hemifield but had little effect on ongoing pursuit of targets in the contralateral hemifield (Basso, Krauzlis, & Wurtz, 2000; Krauzlis et al., 2000; Krauzlis, Basso, & Wurtz, 1997). Stimulation also increased the probability of selecting a stimulus that appeared contralateral to the site of stimulation irrespective of target motion direction (Carello & Krauzlis, 2004). The rSC thus shows two necessary properties for target selection: target/distractor discrimination and forced bias via stimulation. However, stimulation of the rSC does not generate SPEM, and the bias generated by the rSC stimulation is irrespective of the direction of target motion. One possible explanation for these two effects comes from the bilateral connections between the SC and the LIP. The LIP sends collaterals to the intermediate and deep layers of the SC (Lynch, Graybiel, & Lobeck, 1985) and gets return projections from the SC via the pulvinar (Clower, West, Lynch, & Strick, 2001). The LIP shows both these properties for SAC target selection and so, rSC activity might simply reflect the activity of foveal LIP cells during SPEM.

Although the model's SPEM and SAC subsystems each use a frontal–BG–thalamic loop for voluntary gating of target selection, there are some notable differences in the connectivity in the current version of the model. In the SAC system, FEF planning cells, SC buildup cells, and LIP cells all project to the BG–thalamus decision stage and help choose the target. Grossberg et al. (1997) modeled how activity representing multiple possible saccadic targets can coexist during the preparatory phase in the deeper layers of the SC, until competition determines the final choice of saccadic target. Gancarz and Grossberg (1999) further developed these ideas to predict and simulate how the FEF and the SC may work together to determine a saccadic choice. The current model further develops these concepts.

In particular, the FEF motor output cells are distinct from the planning cells, and do not become active enough to drive downstream SAC generators until they receive a GO signal from the BG–thalamus decision stage. In contrast, FPA averaging cells, which are here modeled as the FPA cells that send outputs to downstream SPEM generators, are the sole source of input to the SPEM part of the BG–thalamus decision stage, as well as the sole cortical recipient of the GO signal that initiates high-gain pursuit. Given a low threshold for SPEM initiation downstream from the FPA, the current model cannot avoid exhibiting a phase of low-gain vector-averaging SPEM before the BG–thalamus decision process forces a choice and initiates high-gain pursuit of whichever target direction representation survives the enhanced competition induced by the action of the GO signal.

This lack of full parallelism between the FEF–BG–thalamus and the FPA–BG–thalamus circuits is in accord with data indicating that vector averaging is much more common in the SPEM than in the SAC system. This may be because the visual consequences of SAC vector averaging are highly negative. The same is not true for SPEM vector averaging. In any case, some other variants of the SPEM circuit are also compatible with the data and computational constraints (Figure 6). For example, similar behavior can be achieved if both the MST_v-recipient WTA and averaging cells of the FPA project to the BG–thalamus decision stage, provided that weights and thresholds are adjusted accordingly. What would not work would be a projection from the model FPA's MST_d-recipient summation cells to the BG, if they have a large relative weighting. Such a circuit would exhibit mid-pursuit failures of the GO signal because the discharge rate of the summation cells declines sharply during maintained successful pursuit. This is because the MT cells that drive them lose their major excitatory input when successful pursuit cancels target motion in the retinal frame. In contrast, the MST_v cells (which ultimately drive FPA WTA and averaging cells in the current model) receive an efference copy signal that sustains their discharge during successful pursuit (Pack et al., 2001).

What pathway ensures coordination between the SAC and SPEM systems if their decisions are initially in conflict? The model embodies the hypothesis that the LIP possesses characteristics suitable for acting as a conduit for a multistep transfer of information between the SAC and SPEM systems. Alternatively, it might be proposed that SAC target choice could be transferred to the SPEM system in a single step, via a direct projection from the FEF to the FPA. Although these regions are nearby, no such direct projection is known. How such a projection might be made to work is also unclear because of incongruent representations at the FEF and the FPA. The FEF has a retino-centric vector map, such that each cell explicitly codes a saccade having a particular magnitude and direction and implicitly codes a single retinotopic locus. The FPA, on the other hand, has a directional map, such that each cell explicitly codes a preferred direction of SPEM motion regardless of the target's retinotopic locus. So any direct “alignment” could only be based on directions, not on retinotopic positions. As such, any would-be selective signal from the FEF to the FPA would be ambiguous: There might be many targets moving in the same direction, but at different positions on the retina.

To effect the transfer of choice from SAC to SPEM, the model instead incorporates the well-established projection from the FEF to the LIP to the MT. The transfer can be unambiguous as to retinal locus because the FEF, the LIP, and the SC all embody retino-centric maps. In the model, each region is represented by a 15 × 15 grid with the center [7, 7] representing the fovea. The LIP-to-MT projection can be thought of as a position-based attentional bias signal. This assumption is in tune with data that show MT/MST neurons increasing their activity if a target is present within their retinotopic receptive fields. Enhancement ranges from ∼9%, for random dot coherent motion stimuli (Seidemann & Newsome, 1999), to ∼25–35% for single-stimulus motions (Treue & Maunsell, 1999; Ferrera & Lisberger, 1995). In accord with this, model MT cells, whose receptive fields overlap the retinotopic position of the SAC target choice, get a boost in input. This accords with data that spatial cues reduce latency of eye movement generation more than form or motion cues (Adler et al., 2002). Of course, spatial cueing, by itself, is sometimes insufficient. Because the model has no feature or form representations, in its current form, it cannot simulate how the real SAC system could force a choice, based on color or form, between two moving stimuli falling within the same LIP receptive field.

Once the same stimulus is selected as a target, the SAC and SPEM systems must make further decisions and adjustments to produce the kind of tracking that optimizes visual and cognitive processing of the object. Even in the rare case of linear, constant-speed object motion, target speed can easily exceed the maximum SPEM speed so further decisions include when to generate catch-up saccades. Because catch-up saccades may overshoot, and because jump-back saccades degrade vision, further intelligent adjustments include briefly lowering pursuit gain (the ratio of SPEM velocity to target velocity) below unity, in order to allow the target to more quickly catch-up with the moving line of gaze. A companion paper about the model (Grossberg et al., submitted) treats a large number of such coordination examples and shows a good match between simulation results and experimental data.

Authorship is in rotated alphabetical order. D. B., S. G., and K. S. were supported in part by NSF grant SBE-0354378. S. G. was also supported in part by the Office of Naval Research (ONR N00014-01-1-0624).

Reprint requests should be sent to Stephen Grossberg, Department of Cognitive and Neural Systems, Boston University, 677 Beacon Street, Boston, MA 02215, or via e-mail: steve@bu.edu.