BRAINS

Abstracts Brains:

11.45 – 12.00: Carlos Marques do Rosário (ICE)

Title: Oxide-based resistive switching devices as a basis for brain-inspired information technologies 

Resistive switching (RS) devices are among the most promising candidates for the implementation of neuromorphic computing and other brain-inspired technologies. Nonvolatile RS devices are useful for artificial synapses, while volatile RS devices can be used, for example, to obtain artificial neurons. For both cases, oxides emerge as excellent active materials for the fabrication of RS devices. In particular, vanadium dioxide (VO2), a material that exhibits an insulator-to-metal transition close to room temperature, can be used to obtain volatile resistive switching devices. In this work, planar bridge-structure devices were fabricated from VO2 thin films, which were grown on TiO2 (001) substrates by pulsed laser deposition. The VO2devices show multiple resistance states in a single bridge device, allowing multilevel operation with up to 8 different levels. Beyond that, devices with two identical parallel VO2 bridges with varying bridge-to-bridge distances were fabricated, where multilevel operation was also achieved. This concept can readily be extended to more parallel channels and complex network configurations, offering possibilities for new concepts in brain-inspired computing. 

12.05 – 12.20: Saurabh Soni (HMOE)

Title: Dual-functional memory behavior in EGaIn- and PEDOT:PSS-based devices 

Abstract: Alternative organic electronic memory devices, such as neuromorphic and/or one diode-one resistor (1D-1R) systems can help overcome the limitations faced by conventional semiconductor technologies.[1] Thin films of PEDOT:PSS have been well explored in thermoelectrics, LEDs, solar cells, and recently in neuromorphic devices.[2] Here, we demonstrate the working of PEDOT:PSS device functioning as excellent 1D-1R diodes with abnormally high electrical hysteresis as well as current rectification. Furthermore, we investigated the durability and memory retention of these devices by performing large numbers of WRER cycles. Finally, we aim to fabricate crossbar devices to kill sneak-path currents and show enhanced device functioning.

1. Han, Y., et al. Nat. Mater. 19, 843–848 (2020) 

2. van de Burgt, Y., et al. Nat Electron 1, 386–397 (2018)

12.25 – 12.40: Unai Alegre Ibarra (NE) 

Deep convolutional neural network (DCNN) models are widely adopted for their superior performance and applicability in different areas such as computer vision. Their success comes at the expense of high computational complexity and elevated power consumption costs, associated with the high volumes of data that need to be moved around in the process. Domain-specific hardware (DSH) solutions are the de-facto approach for improving energy efficiency, as they allow to tailor solutions for particular data movement needs. Due to the projected limitations of digital computing, analogue computing components for digital DSH are keenly being sought. We study the potential of dopant network processing units (DNPUs), a novel CMOS-compatible technology, for multiply-and-accumulate (MAC)-less convolution computation. Leveraging this concept, we introduce a general-purpose DCNN engine for low-power computations of convolutions at the inference stage, with an estimated power efficiency of 75 TOPS/W. The core of the engine is applied, in a time-multiplexing fashion, to the MNIST dataset with a variation of the LeNet model (obtaining 97.9% test accuracy), and to an end-to-end self-driving car simulation with a variation of the NVIDIA model, maintaining indefinitely a car on its lane.

12.45 – 13.00: Naoki Kogo (BMS)

A hybrid system to investigate neural dynamics: toward true integration of artificial and biological neurons

We recently developed a system where model neurons and real-life biological neurons interact through simulated synaptic conductance, applying a dynamic clamp approach. We used this hybrid system to investigate a neural competition mechanism that is assumed to underly neural decision-making processes. For this purpose, mutual inhibitory connections were established between two pyramidal neurons in visual cortex (layer 2/3 in V1) by implementing model inhibitory neurons and model synapses between them. In addition, we implemented biologically defined modelled neuronal noises and investigated their effects on neural competition dynamics. To be able to investigate the dynamics of populations of neurons, we are currently developing a system where the dynamic clamp approach is applied to an all-optical system.