Mainstream modeling of neuropsychological phenomena has mainly been focused to reproduce their neural substrate whereas sensorymotor contingencies have attracted less attention. In this study we trained arti cial embodied neural agents equipped with a pan/tilt camera, provided with di erent neural and motor capabilities, to solve a well known neuropsychological test: the cancellation task. Results showed that embodied agents provided with additional motor capabilities (a zooming motor) outperformed simple pan/tilt agents, even those equipped with more complex neural controllers. We concluded that the sole neural computational power cannot explain the (arti cial) cognition which emerged throughout the adaptive process.
Mainstream models of neuropsychological phenomena are mainly based on articial bioinspired neural networks that explain the neural dynamics underlying some neurocognitive functions (see for example [4]). Much less attention has been paid to modeling the structures that allow individuals to interact with their environment, such as the sensory-motor apparatus (see [5] for an exception). The neurally-based approach is based on the assumption that the neural computational power and its organization is the main source of the mental life. Alternatively, as stated by eminent theorists [8, 9, 11], cognition could be viewed as a process that emerges from the interplay between environmental requests and organisms' resources (i.e. neural computational power, sensory-motor apparatus, body features, etc.). In other words, cognition comes from the adaptive history (phylogenetic and/or ontogenetic) in which all living organisms are immersed and take part. This theoretical perspective leads to building up arti cial models that take into account, in embryonic form, neural structures, sensory-motor apparatus, environment structure and adaptation processes (phylogenetic and/or ontogenetic). This modelization approach is developed by the interdisciplinary eld of Artifcial Life and it is widely used in order to modelize a large spectrum of natural phenomena[3, 10, 6, 7]. In this study we applied arti cial life techniques to building up neural-agents able to perform a well known neuropyschological task, the cancellation task, currently used to study the neurocognitive functions related to spatial cognition. Basically, this task is a form of visual search and it is considered as a benchmark to detect spatially-based cognitive de cits such as visual neglect [1]. 2
2.1
The cancellation task is a well known diagnostic test used to detect neuropsychological de cits in human beings. The test material typically consists of a rectangular white sheet which contains randomly scattered visual stimuli. Stimuli may be of two (or more) categories (for example triangles and squares, lines and dots, A and C letters, etc.). Figure 1a shows an example of the task. Subjects are asked to nd and cancel by a pen stroke all the items of a given category (e.g. open circles ). Fundamentally, it is a visual search task where some items are coded as distractors and other represent targets (the items to cancel). Braindamaged patients can fail to cancel targets in a sector of space, typically the left half of the sheet after a lesion in the right hemisphere (visual neglect, see gure 1b).Here we simulated this task through a virtual sheet (a bitmap) in which a set of targets and distractors are randomly drawn (Fig. 1c), and trained neural agents provided with a speci c sensory-motor apparatus, described in the next section, to perform the task.
A neural agent is equipped with a pan/tilt camera provided with a motorized zoom and an actuator able to trigger the cancellation behavior (Fig.2). The camera has a resolution of 350x350 pixels. Two motors allow the camera to explore the visual scene by controlling rotation around x and y axes while a third motor controls the magni cation of the observed scene. Finally, the fourth actuator triggers a cancellation movement that reproduces in a simpli ed fashion the behavior shown by human individuals when asked to solve the task. The behavior of the neural agents is controlled by a neural network able to control the four actuators and to manage the camera visual input. The camera output does not gather all the pixel data, but pre-processes visual information using an arti cial retina made up of 49 receptors (Fig. 3, right). Visual receptors are equally distributed on the surface of the camera; each receptor has a round visual eld with a radius of 25 pixel. The activation of each receptor is computed by averaging the luminance value of the perceived scene (Fig. 3, left) 2.3
In order to simulate a form of cancellation task in silico, we trained neural
agents endowed with di erent neural architectures to perform the cancellation
task. In particular, we presented a set of randomly scattered stimuli made up of
distractors (black stimuli) and targets (grey stimuli) (Fig. 4) and rewarded neural
agents for the ability to nd (by putting the center of their retina over a target
stimulus) and cancel/mark correct stimuli (activating the proper actuator).
In order to perform the cancellation task, an agent has to develop (
In 8 evolutionary experiments, we trained neural agents by varying the controllers' architecture (4 conditions) and by adding the possibility to use or not the zooming actuator (2 conditions). For each experiment 10 populations of arti cial agents were trained through a standard genetic algorithm [8] for 1000 generations. For each generation neural agents were tested 20 times with random patterns of target and distractor stimuli. Each agent was rewarded for its ability to explore the visual scene and correctly cancel/mark target stimuli. 3
For each evolutionary experiment we post-evaluated the best ten individuals for the ability to correctly mark target stimuli. In particular, we tested each individual with 800 di erent random stimuli patterns.The rationale behind the post evaluation is twofold. First, during evolution each agent experienced a small number of possible visual patterns (20); second, the reward function was made up of two parts so as to avoid bootstrapping problems: one component to reward exploration and the second one to reward correct cancellations. Results are reported as proportion correct in cancellation tests. Figure 6 reports the postevaluation results for each architecture in each motor condition: with the ability to operate the zoom (Fig. 6 a,b,c and d) and without this ability (a-, b-, c- and d-).
For all the neural networks we found signi cant di erences (p<0.001, twotailed Mann-Whitney U test) between the condition presence/absence of the capacity to zoom incoming stimuli. In both groups there were signi cant di erences between network a and the remaining networks, but no signi cant di erence emerged between networks b, c and d. Interestingly, there were no signi cant di erences between a, b , c , and d . This last result suggests that a greater computational power can replace to some extent the absence of a zooming capacity. As mentioned above, neglect patients fail to process information coming from the left side of space. However healthy individuals can also show mild signs of spatial bias in the opposite direction (i.e., penalizing the right side of space), a phenomenon termed pseudoneglect [12]. In order to asses if such bias could simply have emerged as a side e ect of the training process, we tested the best evolved individuals of the network d with a set of 200 couples of target stimuli placed symmetrically respect to the x axes of the arti cial agent. Results (Fig. 7) show that only one individual (nr. 3 in Fig. 7) did not present a signi cant left-right di erence, while all the remaining had di erent degrees of spatial preference. At variance with the mainstream approach in the modeling of neuropsychological phenomena, mainly focused on reproduction of the neural underpinnings of cognitive mechanisms, we showed that having a proper motor actuator can greatly improve the performance of evolved neural agents in a cancellation task. In particular, we demonstrated that an appropriate motor actuator (able to implement a sort of attentional/zooming mechanism) can overcome the limits associated with intrinsic computational power (e.g. number of internal neurons and neural connections in our case). Second, we showed that spatial bias in stimulus selection in healthy neural agents can be a side e ect of the training process. In future extensions of this work we plan to test injured neural agents, evaluate biologically-inspired neural architectures following recent research results on brain attentional networks[2] and to extend the range of di erent explored sensory-motor capabilities. 4. Marco Casarotti, Matteo Lisi, Carlo Umilta, and Marco Zorzi. Paying attention through eye movements: A computational investigation of the premotor theory of spatial attention. J. Cognitive Neuroscience, 24(7):1519{1531, 2012. 5. Andrea Di Ferdinando, Domenico Parisi, and Paolo Bartolomeo. Modeling orienting behavior and its disorders with ecological neural networks. Journal of Cognitive Neuroscience, 19(6):1033{1049, 2007. 6. Onofrio Gigliotta and Stefano Nol . Formation of spatial representations in evolving autonomous robots. In Proceedings of the 2007 IEEE Symposium on Arti cial Life (CI-ALife 2007), pages 171{178, Piscataway, NJ, 2007. IEEE Press,. 7. Onofrio Gigliotta, Giovanni Pezzulo, and Stefano Nol . Evolution of a predictive internal model in an embodied and situated agent. Theory in Biosciences, 130(4):259{276, 2011. 8. Stefano Nol and Dario Floreano. Evolutionary Robotics. Mit Press, 2000. 9. Rolf Pfeifer and Josh Bongard. How the body shapes the way we think. The Mit
Press, 2006.
10. Michela Ponticorvo and Orazio Miglino. Encoding geometric and non-geometric
information: a study with evolved agents. Animal Cognition, 13(