Conclusion and Future Work
Amir Zjajo, René van Leuken · River Publishers eBooks · 2022
Current neuron simulators, which are precise enough to simulate neurons in a biophysically meaningful way, are limited in the amount of neurons to be placed on the chip, the interconnect between the neurons, run-time configurability and the re-synthesis of the system. Porting the network to the FPGA yields at least several thousands of simulation speed-ups in comparison with SystemC simulation, with negligible loss of accuracy. Real-time reconfigurable learning neuron networks are not only limited by run-time configurability, the re-synthesis of the system and the interconnect between the neurons, but also mainly by the amount of neurons that can be placed on the chip. The multi-compartment synapse design gives biologically accurate modelling of chemical synapse, which increases the computation ability of synapses. Moreover, the receptors have extremely low energy consumption of only 1.92, 3.36, and 1.11 pJ/spike, representing a good balance between biological complexity and configuration diversity.