Session S41: Neuromorphic Systems: Concepts, Materials and Devices

Spintronic devices for neuromorphic computing

Human brains can solve many problems with orders of magnitude more energy efficiency than traditional computers. As the importance of such problems, like image, voice, and video recognition increases, so does the drive to develop computers that approach the energy efficiency of the brain. Progress must come on many fronts ranging from new algorithms to novel devices that are optimized to function in ways more suited to these algorithms than the digital transistors that have been optimized for the present approaches to computing. Magnetic tunnel junctions have several properties that make them attractive for such applications. They are actively being developed for integration into CMOS integrated circuits to provide non-volatile memory. This development makes it feasible to consider other geometries that have different properties. By changing the shape of the devices, they can be non-volatile binary devices, thermally unstable superparamagnetic binary devices, and non-linear oscillators. In this talk, I describe a few of the computing primitives that have been constructed based on the different functionalities of magnetic tunnel junctions. After a brief overview of other approaches, I will focus on two that I have been involved in. The first of these uses tunnel junctions as non-linear oscillators in the first nanoscale “reservoir” for reservoir computing. The second uses them in their superparamagnetic state as the basis for a population coding scheme. I will discuss the prospects for these primitives to be the basis for energy efficient computing schemes.   

Quantum matter for artificial intelligence and brain sciences

Intelligence in the animal kingdom spans primitive capabilities concerning survival to advanced collective behavior involving co-operation. These field observations serve as inspiration to understand the fundamental mechanisms governing adaptive behavior in nature and further ponder how they may be adapted into the physical world. I will present some biology literature examples of intelligence in non-neural and neural organisms and the underlying features of biological plasticity. I will then discuss mechanisms and realization of electronic plasticity in strongly correlated quantum matter that enable mimicry of several fundamental features of living beings. These properties naturally are suited to implement in machine learning. I will give an example of our recent collaborative study concerning memory loss in electron-doped quantum perovskites and how this property can mimic lifelong learning.
  

Analog Neurocomputing with Emerging Memory Devices

The revolution in deep learning was triggered not by any significant algorithm breakthrough, but by the use of more powerful GPU hardware. Though this revolution has stimulated the development of even more powerful digital circuits, their speed and energy efficiency is still inadequate for more ambitious cognitive tasks. On the other hand, the network performance may be dramatically improved using mixed-signal integrated circuits based on emerging nonvolatile memories, where the key inference-stage operation, the vector-by-matrix multiplication, is implemented on the physical level by utilization of the fundamental physical laws. I will review the recent progress of such mixed-signal neuromorphic networks based on floating-gate memories and metal-oxide memristive arrays. The recent experiment results for 180-nm NOR flash pattern classifier showed >×10³ improvements in propagation delay and energy dissipation per inference operation compared to digital implementation. The performance can be further improved by utilizing memrisistors, which are scalable to 10-nm and suitable for 3D integration. Thier fabrication technology was recently improved to demonstrate the first simple integrated neuromorphic networks.   

Nonlinear dynamics and imaging of current density and electric field bifurcations caused by electronic instabilities

In 1963 Ridley postulated that, under appropriate biasing conditions, a system that exhibits either a current-controlled or a voltage-controlled negative differential resistance will bifurcate, via entropy-production-maximization, to form regions with different current-densities or electric-fields, respectively. The ensuing widespread discussions in the non-equilibrium statistical mechanics community, however, failed to agree on specific mechanisms causing such bifurcations. Using thermal and chemical spectro-microscopy, my group directly imaged current-density- and electric-field-bifurcations in transition metal oxides that are being used to implement threshold resistance switching in memristors and enable new types of neuro-mimetic devices. We found that nonlinear dynamical circuit theory and the principle of local activity successfully predict both chaotic dynamical behavior and current-density- and electric-field-bifurcations, as well as provides a mechanism for why the bifurcations occur. We determined that upon bifurcations, internal enthalpy in the device reduces despite unchanged power input and heat output, thus suggesting an important thermodynamic constraint required to model nonlinear electronic devices. Our results explain the electroforming process that initiates nonvolatile switching in metal oxides and has significant implications for properly modeling any semiconductor device, since bifurcation can occur for many types of activated processes.   

Spiking neural networks with resistive-switching synapses for STDP-based unsupervised learning

Emulating the architecture and learning mode of the human brain is a grand challenge for the modern neuromorphic engineering. Toward this goal, neural networks should feature spiking neurons and plastic synapses capable of high connection density and spike timing dependent plasticity (STDP) for unsupervised learning of external stimuli. This work illustrates the state of the art of spiking neural networks (SNNs) with nanoscale synapses based on resistive switching memory (RRAM) devices. The RRAM device engineering, the synapse circuit design and plasticity, and the hardware demonstration of unsupervised learning in the SNN will be described. The challenges for high-performance SNNs, including the materials engineering for high RRAM reliability, the synapse architecture to enable STDP, the spiking distribution to support unsupervised learning, and the hybrid integration of RRAM synapses and silicon-based neurons on the same chip will be reviewed. Finally, novel research paths for brain- inspired recurrent SNNs featuring associative memory, pattern recognition and error correction will be discussed.