Long-Short Term Memory Networks for Modelling Embodied Mathematical Cognition in Robots

Mathematical competence can endow robots with the necessary capability for abstract and symbolic processing, which is required for higher cognitive functions such as natural language understanding. But, so far, only few attempts have been made to model mathematical cognition in robots. This paper presents an experimental evaluation of the Long- Short Term Memory networks for modeling the simple mathematical operation of single-digits addition in a cognitive robot. To this end, the robotic model creates an association between the proprioceptive information from finger counting and the handwritten digits of the MNIST dataset. In practice, the model executes two tasks concurrently: it recognizes the handwritten digits in a sequence and sums them. The results show that the association with fingers can improve the robot precision, as observed in children. Also, the robot makes a disproportionate number of split-five errors similarly to what observed in studies with children and adults, hence giving evidence to support the hypothesis that these errors are due the use of a five-fingers counting system.