NU 2021-179 PRIMARY INVENTOR Jiangtan Yuan Stephanie Elizabeth Liu Vinod Kumar Sangwan Amit R. Trivedi Mark C. Hersam*
SHORT DESCRIPTION A reconfigurable memtransistor that can dynamically adapt to new learning tasks
BACKGROUND Recent AI advances have been possible due to GPU enabled parallel computing. However, this high computing capacity is not only energy hungry but not well-suited for asynchronous applications. Recent developments in neuromorphic architecture provide one possible solution. Current brain-inspired neuromorphic hardware is a non-volatile low power solution, but it is energy insufficient and cannot adapt real-time to new learning tasks beyond the pre-defined scope. Thus, there is a mismatch in capabilities between current software advances and neural-like intelligent hardware, with the latter being one of the limiting factors for further scientific breakthroughs.
ABSTRACT Neuromorphic devices operate on event-based design and provide an efficient low power solution that is capable of mimicking biological systems. Northwestern researchers have developed a novel neuromorphic device, a multi-terminal and gate-tunable memtransistors that shows promising potential to closely emulate neural connections in an energy efficient manner. Effects of sapphire substrate on the 2D material properties of MoS2 film allow for fine tuning response to voltage inputs that exhibit various distinct learning profiles. Resulting memtransistors are capable of employing spiking neural networks to dynamically learn as well as learn to forget, and to do so continuously as well as unsupervised. This ability to reconfigure their application is a unique demonstration of a non-volatile solid-state electronic device that can dynamically adapt to learning new tasks by repurposing synaptic units over the device’s lifetime, closely mimicking biological neural connection plasticity and offering a platform for new integrated AI applications.
APPLICATIONS
ADVANTAGES
PUBLICATIONS Yuan J, Liu SE, Shylendra A, Gaviria Rojas WA, Guo S, Bergeron H, Li S, Lee H, Nasrin S, Sangwan VK, Trivedi AR, and Hersam MC (2021) Reconfigurable MoS2 Memtransistors for Continuous Learning in Spiking Neural Networks. Nano Letters 21. 15:6432-6440. IP STATUS Provisional application has been filed.