“…At reverse, knowing the variation in the sensory maps, it is possible to estimate which transformation (hidden variable) is the most probable to have generated these outputs. This property of auto-encoders can serve for active inference and action observation, which are also features observed in parietal neurons and in the mirror neurons system [28], [29], [30] for affordances generatation [31], [32] and also sensorimotor adaptations as during tool-use [18]. During grasping, the prediction done by the motor units of the hidden layer of the auto-encoder can serve to "reverseengineer" the hand preshaping based on visual information; this idea is also found in Rumelhart or Kawato's forwardinverse models [33], [34] as well as in the "virtual finger hypothesis" by Arbib who proposed to explain grasp affordance and the assignement of the orientation and of the power grip of the real fingers during grasping [31], [32], [35].…”