Pseudoneglect in Visual Search: Behavioral Evidence and Connectional Constraints in Simulated Neural Circuitry

Discussion

In this study, we established specific connectivity constraints leading to a lateral spatial bias (pseudoneglect) in artificial organisms trained to perform a visual search task by using genetic algorithms. A form of pseudoneglect that was qualitatively and quantitatively similar to that shown by normal participants did emerge in artificial neurorobots, but only in those harboring hemispheric asymmetries of connectivity that simulated those typically occurring in the human brain. As a further condition, a general excitatory influence of VAN on the ipsilateral DAN was necessary for pseudoneglect to occur in neurorobots. This novel result suggests that hemispheric asymmetry alone is not sufficient to generate a leftward bias, and thus further specifies the likely connectional constraints of pseudoneglect.

We first consider our results in the light of neurophysiological studies of pseudoneglect, and then in relation to existing modeling studies of the human attentional system. A particular instance of pseudoneglect occurs with the landmark task: When judging lines prebisected to the left of their true center, normal participants consider the left segment as being longer than the right one (Milner et al., 1992). Spatial attention has been shown to be a major determinant of this phenomenon (Toba et al., 2011). Szczepanski et al. (2010) and Szczepanski and Kastner (2013) tested normal participants’ spatial bias on covert attention tasks and on the landmark task by using a multimodal approach, combining psychophysics, fMRI and transcranial magnetic stimulation (TMS). They tested only frontal and parietal ROIs in the DAN, and did not explore the VAN. Their subjects’ sample showed a mixed spatial bias: some subjects had a leftward bias (pseudoneglect), but most subjects showed a rightward bias (Szczepanski and Kastner, 2013). On average, the bias was rightward, unlike most of the literature results. The lateralization of the bias correlated with the lateralization index of the fMRI activation in the ensemble of the DAN ROIs during a covert spatial attention task. Specifically, subjects that had more left hemisphere activation also had a contralateral, i.e., rightward, bias in the landmark task; conversely, subjects with more right hemisphere activation tended to have a leftward behavioral bias. TMS-induced interference on the left- or right-hemisphere parietal nodes during the landmark task caused an ipsilateral shift of the bias: right parietal TMS caused a rightward shift compared to the initial bias, and left parietal stimulation caused a leftward shift. Stimulating both right and left parietal ROIs did not cause a shift. Szczepanski and Kastner (2013) suggested that there is an interhemispheric competition between the DAN nodes, and the lateralization of the sum of the weights in the DAN activation shifts the attentional focus contralaterally. The possibility of long-range suppression, which might involve the DANs in both hemispheres, was shown in the monkey LIP: firing rate was suppressed when a saccade target was as far as 50° from the neuron receptive field (Falkner et al., 2013).

Thus, these results are broadly consistent with the functioning of the present neurorobot population C. In agreement with Szczepanski and Kastner (2013)’s conclusions, the DAN in the current model is conceptualized as a whole, and not as separated nodes. Additionally, Szczepanski and Kastner’s data showed that there is large variability between participants in the direction and degree of lateralization of DAN activation, that on average did not significantly differ between the hemispheres. Here, we aimed to explore the typical functional architecture in the human population. Therefore, we chose to model the DAN as laterally symmetrical and the VAN as right lateralized. However, there are several differences between the current models and Szczepanski et al. (2010). First, they used a landmark task while here we used a search task. Second, the overall behavioral pattern here was of a leftward classical pseudoneglect bias and not the rightward bias found by Szczepanski et al. (2010), This might result from substantial differences in the studied samples or in the tasks used. Third, and more importantly, the VAN, which has a major contribution in the current model, was not tested in their studies.

The architecture of neurorobot C is partly inspired by the results of Koch et al. (2011), which might oversimplify the nature of interhemispheric interactions. Several fMRI studies of human attention areas found evidence of bilateral activation of attention areas, with a contralateral bias (Patel et al., 2015). In neurorobots D and E, we introduced bilateral competence in the right hemisphere networks (Heilman and Van Den Abell, 1980; Mesulam, 1981). However, performance in these models showed no consistent spatial bias. This suggests that right hemisphere bilateral competence by itself might not be crucial to the emergence of pseudoneglect. Moreover, in our setting the inhibitory valence of inter-DAN connections was important for the development of an initial leftwards spatial bias, as well as to reach optimal levels of performance, as stressed by additional control simulations in which inter-DAN connections were either set free to evolve as inhibitory or excitatory (neurorobots F), or constrained to assume only excitatory valence (neurorobots G). On the other hand, evidence from neglect patients (Bartolomeo and Chokron, 1999) challenges models of attention exclusively based on interhemispheric rivalry (Kinsbourne, 1970, 1977, 1993). In addition, bilateral competence in attentional areas might be important in long-term compensation of neglect (Lunven et al., 2015; Bartolomeo and Thiebaut de Schotten, 2016). Our results stressing the importance for pseudoneglect of both right-hemisphere bilateral competence and interhemispheric competition may thus pave the way for an integrated interpretation of different lines of research on normal or dysfunctional human attention networks.

In their recent review, Borji and Itti (2013) provided a taxonomy of nearly 65 computational models of visual attention. Many of these models focused on reproducing eye movements (e.g., the saliency-based models reported in Borji and Itti, 2013), following a bottom up approach. Typically, these models extract a set of features, represented as maps, from an incoming image. Then, feature maps are combined in a saliency map where a winner-take-all mechanism will designate the spatial region to be attended. Saliency-based attention models in general do not account for exploration biases, with the exception of a recent model (Borji and Tanner, 2016), where an object center bias (the tendency to focus on the center of objects) is reproduced by adding an ad-hoc bias map to the saliency map. While important for building predictive models, this result seems of little relevance to lateral biases such as pseudoneglect. Other models (Deco and Rolls, 2004; Deco and Zihl, 2004) simulated attention as emerging from the competition of several brain areas subjected to bottom-up and top-down biases. These models do not drive eye movements; the scan path is simulated as a sequence of activations of the simulated posterior parietal cortex. Lanyon and Denham (2004, 2010) added to these models simulated eye movements and an adjustable attention window scaled according to stimuli density. Despite being successful at reproducing scan paths in healthy individuals and neglect patients, these models do not address the issue of pseudoneglect. Other models of attention did not consider pseudoneglect because of their training procedure or design constraints (Pouget and Sejnowski, 2001; Mozer, 2002; Monaghan and Shillcock, 2004; Di Ferdinando et al., 2007). Di Ferdinando et al. (2005) explored line bisection and target cancellation performance in four biologically inspired neural networks. The networks’ patterns of connectivity varied along different degrees of asymmetry, inspired by specific theories. Pseudoneglect occurred in line bisection but not in visual search. In these models, motor outputs were only used for target selection; there was no active exploration of the environment, whereas when our neurorobots explored their environment the corresponding input information changed as a function of eye movements. Nonetheless, the present study shares with Di Ferdinando et al. (2005) and other work from the Zorzi group (Casarotti et al., 2012) the stress on accounts of attentional phenomena relying on sensory-motor transformations, as stated by the premotor theory of attention (Rizzolatti et al., 1987). Specifically, our results support the hypothesis that the way in which the movements of the actuators are controlled affects the performance on a cancellation task (Gigliotta et al., 2015).

Thus, contrary to most available models of attention, our artificial robots are trained to correctly cancel target stimuli, and are free to self-organize to find a proper solution, within the sole limits of the imposed connectivity constraints. These constraints were inspired by available data concerning the anatomic and functional organization of the attentional networks in the human brain. To the best of our knowledge, this is the first attempt to simulate pseudoneglect as a consequence of activity in the dorsal and ventral attention networks in the two hemispheres of the human brain. While this article was under review, two theoretical papers were published that also took into account the dorsal/ventral architecture of the attentional networks (Parr and Friston, 2017; Seidel Malkinson and Bartolomeo, 2017), but neither endeavored to simulate pseudoneglect. Another original feature of the present models is the embodiment factor, consisting of the explicit modeling of eye movements (Bartolomeo et al., 2002; Lanyon and Denham, 2004; Di Ferdinando et al., 2007; Miglino et al., 2009; Gigliotta et al., 2015). In particular, the present models extended the models devised by Di Ferdinando et al. (2007), by increasing the complexity of the organisms’ retina, the biological plausibility of the motor system and that of the neural controllers. Conti et al. (2016) also adopted an embodied perspective, based on a humanoid robot platform. In their study, an iCub robot was trained to remove objects from a table, a task reminiscent of a cancellation task. Intrahemispheric disconnections were able to produce neglect-like behavior. However, the embodiment of the model was limited by the facts that selection of a visual target was conducted independently of the motor behavior and that the robot’s eyes were kept fixed during the cancellation task. Moreover, although hemisphere asymmetry was modeled by increasing the number of right hemisphere processing units, no bias in normal performance is reported.

Moreover, contrary to most published work, our model attempted to simulate the relationships between the visual pathways and the attentional networks by respecting important biological constraints. Visual pathways project mainly to the hemisphere contralateral to each visual field. However, theoretical models of visual attention posit that the left hemisphere mainly deals with the contralateral hemispace, whereas the right hemisphere has a more bilateral competence (Heilman and Van Den Abell, 1980; Mesulam, 1981). In previous computational models this asymmetry has not always been simulated in a biologically plausible way. In some cases, both simulated hemispheres received visual information from the whole visual field, with attention asymmetries being represented in inner layers (Monaghan and Shillcock, 2004; Di Ferdinando et al., 2007). In Conti et al. (2016)’s model, the right hemisphere received information from both visual hemifields, whereas the left hemisphere processed only the contralateral visual hemifield. Our models D and E had similar architecture, but were unable to mimic human performance. Moreover, there is no anatomic evidence of such asymmetries in the visual pathways, and information exchange in the occipital visual areas is mainly limited to the vertical meridian (Berlucchi, 2014). In our model, these important biological constraints of visual information processing were respected, because each artificial hemisphere received visual information from the contralateral hemifield; interhemispheric connections were only present at a later stage of processing, between the artificial DANs.

It might be argued that in our model C a leftward bias was simply transferred or amplified from the input to the output layers. If so, however, we would have expected to observe a constant leftward bias, akin to right-sided neglect. What we found, instead, was just an initial leftward bias, at the onset of the exploration task, analogous to human physiologic pseudoneglect. To observe this initial bias, the VAN-DAN connections had to have an excitatory valence. This occurrence does not result from existing empirical data and is thus a novel prediction of the model. Also, neurorobot populations D and E, which also had more right hemisphere than left hemisphere resources, and should then have a similar input-to-output amplification, did not show pseudoneglect, presumably because of the lack of interhemispheric inhibition.

The level of detail of the models is not a trivial matter, because it has to provide meaningful novel information while remaining tractable. A potential limitation of our study is the use of simplified versions of the fronto-parietal cortical networks, without taking into consideration the substructures of the DAN and VAN, which are both broad and partly heterogeneous networks (Colby and Goldberg, 1999), nor subcortical structures such as striatum, thalamus and superior colliculus (Krauzlis et al., 2017). For example, the connectional anatomy of VAN components such as the temporoparietal junction (e.g., with the ventral cortical visual stream) and of the ventrolateral prefrontal cortex (e.g., with limbic structures) is likely to be crucial to the functioning of the VAN. Yet, our simplified model, with a VAN receiving visual input and sending excitatory connections to the ipsilateral DAN, was able to mimic human performance to an impressive level of accuracy.

More generally, our modeling is consistent with evidence from healthy subjects and neglect patients, stressing the importance of entire fronto-parietal networks, or of their dysfunction, in behavioral patterns such as pseudoneglect (Szczepanski and Kastner, 2013), or visual neglect (Corbetta and Shulman, 2011; Bartolomeo et al., 2012), respectively. Also, integrated fronto-parietal activity, with subtle, task-dependent differences in network dynamics, occurs during attention orienting in monkeys (Buschman and Miller, 2007). Concerning visual neglect, evidence suggests that a major determinant of this condition is indeed a dysfunction of the right hemisphere VAN (Corbetta and Shulman, 2011; Urbanski et al., 2011), or of its connections with the ipsilateral DAN (Thiebaut de Schotten et al., 2005).

Finally, we note that the present population-based model can be potentially used to explore in a natural manner the universal properties (the basic brain architecture) and individual differences in network efficiency, two aspects recently underlined by Posner (2014) as appropriate features for future models of attention.

In conclusion, we have demonstrated the emergence of pseudoneglect behavior in artificially evolving neurorobots searching for visual objects, under specific connectional constraints. These neurorobots provide a plausible model for the dynamic interactions between fronto-parietal attention networks in the human brain.