Neural networks - based solutions for predicting the affective state of children with autism

Autism spectrum disorder (ASD), is a neurobehavioural disorder, identified by difficulties in speech, communication, different behaviour and it cannot be cured entirely [21]. This paper presents a comparison between Feedforward Neural Networks and Convolutional Neural Networks, using a children drawings dataset, created from the ground up, in order to train the best model. The drawings dataset is divided in five groups, each corresponding to an affective state: happy, fear, sad, insecure and angry. The model will be used in a mobile application to predict the affective state of the children, with ages between 2-4 years. Firstly, in this paper is set out a parallel between VGG16 and MobileNet, from which the results concluded, that the MobileNet model performed better, obtaining an accuracy of 58%. Is continued with the use of Feedforward Neural Networks, exposing the following results: loss=1.5597 and accuracy= 28.3333%.