Study of Spiking Hebbian Memory Networks for implementation on the Neuromorphic chip LoihiPREMK, ALEN (Avtor) Kranjc, Matej (Mentor) Legenstein, Robert (Komentor) Hebbian plasticityNeuromorphicLoihiSpiking Neural NetworksIntroduction: The underlying mechanisms in the workings of memory in the human brain are still largely unknown. Longer-term storage capabilities that are based on Hebbian synaptic plasticity can be tested on artificial Hebbian memory networks. Neuromorphic hardware can be used for employing SNN forms of algorithms, which can provide further insights to how memory is formed and retrieved in the biological brain. The aim was to construct a simple spiking Hebbian memory network model algorithm which could further be implemented to Intel’s neuromorphic chip Loihi. Methods: Non-spiking and spiking Hebbian memory networks were described and tested. A synthetic dataset which represents a simplified version of the bAbI dataset was created. Hebbian network models were modified accordingly to be able to solve the synthetic dataset task. The experiments were performed on the networks solving the bAbI task 4 and the synthetic dataset task. Results: Non-spiking H-mem and spiking H-mem network models were able to solve the bAbI task 4 and the synthetic dataset task. In experimental cases when optimal parameters were chosen, the networks were able to achieve around 100% accuracy when predicting an answer. Conclusions: Good learning accuracy was achieved when experimenting with different Hebbian memory networks. A simple proof of concept was designed and tested with good results. The proof of concept can be used as a basis for implementation of the algorithm to the neuromorphic chip Loihi.20212021-10-19 09:30:05Magistrsko delo/naloga132241VisID: 58401COBISS_ID: 81613059sl