Close
Variance in Memristors: Mitigation and Exploitation
26 February 2021, 10:00 am–11:00 am
Memristors are playing a pivotal role in accelerating machine learning hardware. In this seminar we discuss this how variances between devices can be used to carry out computations when attempting to replicate or interface with bio-inspired computers. But what if we want to avoid these variances? Well we’ll cover that too, discussing the benefits of committee machines in mitigating for device variances. Speakers include Dr. Melika Payvand from ETH-Zurich and Dovydas Joksas from UCL EEE.
This event is free.
Event Information
Open to
- All
Availability
- Yes
Cost
- Free
Organiser
-
Daniel Mannion
Embracing Non-Idealities in Neuromorphic Computation Speaker: Dr. Melika Payvand (INI ETH Zurich)
The multi-state and non-volatile properties of emerging memory technologies, aka resistive memory, holds a lot of potential for the future of neuromorphic technologies. However, the fundamental resistive switching mechanism introduces stochasticity, non-linearity and drift. These properties are considered “non-ideal” for the conventional computing schemes. In this short talk, I will talk about how to embrace these properties for unconventional neuromorphic computation.
Memristive Neural Networks Work Better in Teams Speaker: Dovydas Joksas
Memristor-based implementations of neural networks have been a popular area of research for at least a decade now. Although these physical realisations make neural networks both faster and more energy-efficient, memristor non-idealities can significantly decrease their accuracy. While the traditional way to deal with this had been to mitigate the individual non-idealities, we took a different, system-level approach. We showed that by combining non-ideal physically implemented neural networks into teams and averaging their outputs, we could achieve higher accuracy without sacrificing other desirable properties.