Neuromorphic accelerators for computing
Rodolphe Héliot (Research scientist at CEA-LETI, Saclay, France)
Abstract: Because of power and reliability issues, computer architects are forced to explore new types of architectures, such as heterogeneous systems embedding hardware accelerators. Neuromorphic systems are good candidate accelerators that can perform efficient and robust computing for certain classes of applications. We develop a spiking neurons based accelerator, with its hardware and software that can execute a wide range of signal processing applications. A library of operators is built, and automated place-and-route tools map the application onto the hardware. A mixed-signal, 65nm integrated circuit implementation is designed that focuses on fault tolerance and low-power. Altogether, this system aims at providing to the user a simple, power-efficient way to efficiently implement signal processing tasks on neuromorphic hardware.
Memristor based neuromorphic architectures: from concept to reality.
Abstract : Neuromorphic architectures are frequently considered as ideal candidates to exploit emerging nano-devices that differ radically in their functionality from MOS transistors. Especially, their memristive behaviour is often compared to synaptic connexion in neural network and authors rely frequently on the supposed intrinsic robustness of neural architectures to overcome the defect of nanodevices. Thereby learning procedures have been proposed for ideal memristors. Nevertheless, the actual characteristics of memristors are far from ideal and the supposed robustness of neural networks is not systematically guarantied. Thus, new strategies should be explored to move from
concept of memristor based neuromorphic architecture to real and functional circuits.