Title: Memristor-CMOS hybrid circuits and systems for brain-inspired computing

Presenters:

Kyeong-Sik Min, Kookmin University, Korea

Fernando Corinto, Politecnico di Torino, Italy 

Abstract: 

Due to the rapid advancement of software-based deep learning technology, artificial intelligence (AI) starts to give deep and huge impact on human life recently. The machines which are trained with huge amounts of data-set and energy can outperform humans in some applications such as speech recognition, image recognition, etc. However, if we compare the energy efficiency and data efficiency between humans and machines, humans are still much better than the machines. Human brain is known consuming only as small as ~20W, while the machines need energy with 3 or 4 orders of magnitude higher for performing the same level of cognition and intelligence with the brain. For the data efficiency in training, humans can learn to recognize some items such as cat very easily, while the machines need more than 10 million images for the training process. Hence, for the next-generation AI technology, we should focus on how we can reduce this huge gap in data and energy efficiency between humans and machines. 

In this tutorial, we suggest memristor-CMOS hybrid circuit as a viable candidate for the next-generation AI hardware and explain why the hybrid circuit is better than the conventional CMOS circuit in terms of energy efficient and data-efficient AI hardware. Memristors which were derived from the concept of memory plus resistor can perform both computation and memory at the same site. This unification of computation and memory in memristors can eliminate the old problem of inefficient classical Von-Neumann architecture, which is caused from the fact that the computation is separated from the memory. This separation constantly demands a huge amount of power in interfacing between the memory unit and computation unit. In addition, the complex integration of information performed by biological neural systems is based on several dynamical mechanisms. Among them, the most worth is the synchronization of neural activity. Synchronization of neural activity is also one of the proposed solutions to a widely discussed question in neuroscience: the binding problem, i.e. how our brain binds all the different data together to recognize objects. The natural non-linear dynamics of memristors can be useful in realizing brain-mimicking information processing which is based on non-linear relationship of neurons and synapses.