Motivation

We aim at investigating the neural mechanisms of visuo-motor imitation through convergent evidence from neuroscience, such as neurophysiology, neuroimaging and brain lesion studies. In particular, we consider apraxia, a deficit in imitation following callosal or parietal brain lesion. Our approach is based on the rational that looking at how imitation is impaired can unveil its underlying neural principles.

We have developed a neurocomputational model of visuo-motor imitation, based on self-organizing maps whose neural activities are associated through anti-hebbian learning. We ground the functional architecture and information flow of our model in brain imaging studies, whereas findings from monkey brain neurophysiological studies drive the choice of implementation of our processing modules. The patterns of impairment of our model account for the scores found in a clinical examination of imitation, described below.

Fig. 1. Goldenberg's seminal study of imitation (from Goldenberg et al., Neuropsychologia, 2001).

Goldenberg's seminal study of imitation of meaningless gestures

A seminal study of imitation of meaningless gestures examines a patient with callosal brain lesion, i.e., disconnected brain hemispheres (Goldenberg et al., Neuropsychologia, 2001). The patient was asked to imitate a set of visual stimuli that present different positions of the hand relative to the head. To disentangle the contribution of each hemisphere the patient was tested tachistoscopically (i. e., the stimulus was presented either to the left or right visual field) in a left- or right-hand imitation condition. As shown on Fig. 1 the pattern of errors varies as a function of the visual field to which the stimuli were displayed and the hand used to execute the imitative movement.

Simple information flow model

We suggest a model based on hypothesized non-uniform flow of information through specific regions in the two brain hemispheres, inspired from brain imaging and brain lesion studies. This model explains the difference of performance observed in the four conditions in Goldenberg's study. A schema of the model is shown below: the stimulus is visually processed in the hemisphere contralateral to the visual field (due to optic chiasm) and the motor command is prepared in the hemisphere contralateral to the hand. The arrows show the necessary transfer of information between the two hemispheres, thus a possible source of spatial errors in the imitation (remember that the patient suffers from disconnected hemispheres). Imitation is perfect only in the right visual field - right hand condition, indicating a lateralization of the processing to the left hemisphere and a necessary computational process in the brain area shown in dark grey.

Fig. 2. Schema of hypothesized information flow across both brain hemispheres in the four conditions in Goldenberg's study.

Neurocomputational model of imitation

Our neurocomputational model of imitation (shown in Fig. 3) is composed of three neural networks, a face visual network in Brodmann Area BA 19/37 at the level of the occipito-temporal junction, a face somatic network in area BA 40 in the parietal cortex and a hand position network probably in dorsal premotor area BA 6. The model implements a visuo-motor route mediating somatic knowledge of the human body, as it appears to be the case in imitation of meaningless gestures. The face visual network receives geometrical properties of the visual stimulus to imitate (such as the position and angle of the hand relative to the nose). The face somatic network receives input from the face visual network and somatic input from tactile sensors of the face. The hand position network receives visuo-somatic input from the face somatic network and proprioceptive input from the arm. The neurons in our model are leaky integrator neurons in order to account for variations of the membrane potential in time and have integrating properties.

Simulation of the lesion

We train the neurons' weights with unsupervised algorithms (Kohonen's algorithm and antihebbian learning) such that the networks' neural activity is topologically consistent and has learned the associations between the position of the arm, the tactile stimulus it produces on the face if touched and its visual perception. In the end, visual presentation of the stimulus to imitate alone yields the corresponding neural activities in the face somatic and position networks thus guiding a correct imitative action. For simulating the lesion of the corpus callosum (i. e., impaired transfer of information across the two hemispheres) we have taken into account two observations. First, some of the visual information must cross the callosum since the patient succeeds to imitate some hand positions when he/she visually processes the stimulus in one hemisphere and prepares the motor command in the other hemisphere. Second, if the patient is given "unlimited time" he/she imitated correctly. So we introduce a probability of information transfer either at the level of the connection or at the level of the input of the neuron and we use an integrating factor greater than the decay factor. For a snapshot of the simulation see Fig. 4.

Results

We replicated the experimental conditions from Goldenberg's study of imitation of meaningless gestures: we used the same visual stimuli and same stimulus presentation time, the weights impairment was coherent with the four conditions described in the study. Our model could very well account for the impairement scores observed (see Fig.5, left).

Predictions of the model

The representation of the face in the brain is non-uniform, some face parts are more represented (i.e. eye and mouth) as compared to others (i.e. chin). Therefore we observe inhomogeneities in the precision and processing times of the imitation task dependent on the final position (see Fig.5, right). Focal lesions, i.e., small and localized lesions of our model, confine the imitation impairement to some parts of the face only, or to spatial errors shifted along the coordinate axes used. Finally, severe lesions correlate with longer processing times. We suggest that the time needed to do a correct imitation should be used as a measure of the severity of apraxia, which is rarely correlated to the size of the lesion). Stroke and callosal lesion data should be used to validate the model, whereas the learning properties of our model could account for some of the brain reorganization observed following stroke. For this reason, we have been conducting neuropsychological experiments of imitative apraxia in collaboration with the Geneva University Hospital (HUG) and Centre Hospitalier Universitaire Vaudois (CHUV). For more details on this study please refer to our Experimental study of Apraxia.

Fig.5 Left,comparison of the results of Goldenberg’s study (light grey histograms) to the results of the simulations using the two impairment models: 1) the integrated neuron input is impaired independently, in dark grey, the parameters found are &lambda = 0.5 (integration factor) and &rho = 0.3 (impairment probability); and 2) each weight is impaired independently, in black, parameters found &lambda= 0.4 and &rho = 0.4; the decay factor &tau was set to 30ms in both models. The imitation was considered correct if the error distance was lower than a certain threshold. Right, inhomogeneities in the precision and processing time of imitation gestures toward different parts of the face, dependent on how well represented they are in the face somatic network (in our case the eye has a larger representation than the chin).

Fig.3 Schema of the neurocomputational model of imitation.

Fig. 4. Snapshot of the simulation of the lesion, a visual and tactile stimulus on the face (left) and the corresponding networks' neural activity.

Extension of the model

Fig. 6. Simulation of the tactile surface of the human body used for neurally learning the entire body forward kinematics.