A Deep Learning Neural Network for Number Cognition: A bi-cultural study with the iCub

Abstract

The novel deep learning paradigm offers a highly biologically plausible way to train neural network architectures with many layers, inspired by the hierarchical organization of the human brain. Indeed, deep learning gives a new dimension to research modeling human cognitive behaviors, and provides new opportunities for applications in cognitive robotics. In this paper, we present a novel deep neural network architecture for number cognition by means of finger counting and number words. The architecture is composed of 5 layers and is designed in a way that allows it to learn numbers from one to ten by associating the sensory inputs (motor and auditory) coming from the iCub humanoid robotic platform. The architecture performance is validated and tested in two developmental experiments. In the first experiment, standard backpropagation is compared with a deep learning approach, in which weights and biases are pre-trained by means of a greedy algorithm and then refined with backpropagation. In the second experiment, six bi-cultural number learning conditions are compared to explore the impact of different languages (for number words) and finger counting strategies. The developmental experiments confirm the validity of the model and the increase in efficiency given by the deep learning approach. Results of the bi-cultural study are presented and discussed with respect to the neuro-psychological literature and implications of the results for learning situations are briefly outlined.