Neuromorphic Elements
A. Embedded, Non-volatile Analogue Memory
Introduction :One main challenge in realising neural models in VLSI lies in the implementation of various non-linear functions with only one type of transistor, the MOS, in the standard CMOS process. In addition, an efficient way of storing and updating a massive number of parameters on-chip is also important to gift neural VLSI models with adaptability. This project explore CMOS-compatible devices for implementing non-linear functions and non-volatile memories of neural VLSI models. In addition to storing analogue parameters, non-volatile memories such as RRAM are also applied to neuromorphic circuits that mimic the synaptic plasticity of biological neurons.
Characteristics :Based on knowledge in device physics and VLSI layout technique, devices with customised characteristics will be designed, tested and modelled. Circuits consisted of the customised devices will then be designed and tested.
Pre-requisites:Knowledge in Electronics and Semiconductor Physics.
Professional skills obtained:Experience in device design and device modelling. Neuromorphic design with customised CMOS devices.
Fig.1: (a)Current-mode non-volatile analogue memory circuit and (b)its chip photo (c) the measured programming dynamics as the memory is programmed to different current levels. (d) The voltage-mode analogue memory (e) The measured programming dynamics as the memory is programmed to different voltage levels (f) Bidirectional programming of the analogue memory.
B. Noisy-adaptable Transistors
Introduction :Our research in stochastic neural computation indicates that it would be possible to exploit the intrinsic noise of transistors for useful computation. This novel idea is particularly attractive as the noise and mismatches of advanced semiconductor technologies become non-negligible. In response to this issue, we have devised two types of noise-adaptable transistors, which are compatible with the standard CMOS technology. One is called the octagonal, dual-gate field-effect transistor (ODGFET), and the other called the resist-protective-oxide field effect transistor (RPOFET). The noise levels of these devices can be enhanced by more than two orders than conventional devices. In addition, the noise levels are adaptable through the STI gate voltage for ODGFET and the drain voltage for the RPOFET.
Fig.2 (a)ODGFET and its noise spectra adapted by the STI gate voltage (Vx) (b) RPOFET and its noise spectra adapted by the drain voltage (VD)
