A SPIKING NEURAL NETWORK BASED APPROACH FOR BINAURAL SOUND LOCALIZATION

In recent years, a new version of highly brain-inspired Neural Networks was developed. The Spiking Neural Networks are considered the third-generation of NN trying to reproduce the natural computing and behaviour of the brain; especially in the sensory field, hardware and software architectures have been developed to mimic the biological-neural organization with new electronic technologies to a new direction for the NN approach. The SNN are inserted in the field of computational neuroscience for the modelling of specific brain areas, in natural or pathological behaviour, with a wide variety of neuron models; in the biomedical field, especially in the neuroprostheses, the neuromorphic engineering have started to play an important role to realize energy efficient and real-time capable devices. On the other hand, SNN has become integral part of Artificial Intelligence field to be applied in tasks such as classification, pattern recognition and clustering. There is evidence in literature that the performances of these networks are very promising. The aim of this thesis is the evaluation of Spiking Neural Networks approach performance applied for binaural sound source localization in the space as locator and features extractor, using a model of the human peripheral auditory pathway from the literature. Classic machine learning approaches (SVM,KNN) were applied on the extracted features to improve the results of SNN localization and compute a comparison in sound localization performances. The SNN model for the sound localization reproduces the human peripheral auditory pathway and it was implemented following a model already known in literature. The implementation mimics the entering of the sound in the ears until the arrival in sub-cortical organs to compute the localization in a unsupervised manner through the output layer with the Winner-Takes-All approach, in terms of azimuth and elevation angles.The evaluation of the localization performance was computed based on the firing rate correctness. The extracted features in output contains frequency and location specific information to be employed to train classical machine learning networks with the purpose to overcome the classification difficulties of SNN, improving the results and evaluating the performances in the classification of unseen sounds. The sounds to feed the network are virtually created from different azimuth and elevation points in the space, in anechoic environment with a constant radius distance. The results show that the implementation of the peripheral auditory model reports similar performance compared to the original work, with a greater error in the localization of the pure tones (for the azimuth). The better results were obtained in white noise localization while vowels are coherent with the results of pure tones. The SVM and KNN are able to improve and provide better results in the binaural localization (SVM linear: 99.6%; SVM rbf: 100%); although the single neurons firing rate provides optimal results, the neuron assemblies are able to overcome the limitation of SNN (SVM linear - KNN: 97.86%, SVM rbf: 97.32%) with less time consuming. The accuracy of SVM and KNN reflects the goodness of SNN to extract frequency and location information. The cooperation between SNN and SVM-KNN is very promising in the field of binaural localization, although future works are needed to investigate deeply the potentiality of SNN based approach.