Modeling Self-organized Emergence of Perspective In/variant Mirror Neurons in a Robotic System

A major role attributed to mirror neurons, according to the direct matching hypothesis, is to mediate the link between an observed action and agent's own motor repertoire, to provide understanding “from inside”. The mirror neurons gave rise to various models but one of the issues not tackled by them is the perspective in/variance. Neurons in STS visual areas can be either perspective selective or invariant and the same variability was later also discovered in premotor F5 area in macaques, showing the existence of different types of mirror neurons regarding their perspective selectivity. We model this as an emergent phenomenom using the data from the simulated iCub robot, that learns to reach for objects with three types of grasp. The neural network model learns in two phases. First, the motor (F5) and visual (STS) modules are trained in parallel to self-organize modal maps using the corresponding data sequences from the self-perspective. Then, F5 area is retrained using the output from the pretrained STS module, to acquire the mirroring property. Using the optimized model hyperparameters found by grid search, we show that our model fits very well empirical observations, by showing how neurons with various degrees of perspective selectivity emerge in the F5 map.