Memristor-based Bayesian neural network

Researchers from CEA-Leti, CEA-List, and CNRS built a complete memristor-based Bayesian neural network implementation for classifying types of arrhythmia recordings with precise aleatoric and epistemic uncertainty.

While Bayesian neural networks are useful for at sensory processing applications based on a small amount of noisy input data because they provide predictive uncertainty assessment, the probabilistic nature means increased energy and computation requirements from the use of random number generators, which store the probability distributions.

“We exploited the intrinsic variability of memristors to store these probability distributions, instead of using random number generators,” said Elisa Vianello, CEA-Leti chief scientist, in a release. The approach to performing inference requires massive parallel multiply-and-accumulate (MAC) operations. “These operations are power-hungry when carried out on CMOS-based ASICs and field-programmable gate arrays, due to the shuttling of data between processor and memory. In our solution, we use crossbars of memristors that naturally implement the multiplication between the input voltage and the probabilistic synaptic weight through Ohm’s law, and the accumulation through Kirchhoff’s current law, to significantly lower power consumption.”

The approach enables uncertainty quantification, which allows the network to identify situations that could lie outside of its training data. [1]

Hybrid phase-change memristors

Scientists from the University of Rochester developed hybrid resistive switches that combine memristors and phase-change materials.

“We’ve combined the idea of a memristor and a phase-change device in a way that can go beyond the limitations of either device,” said Stephen M. Wu, an assistant professor of electrical and computer engineering and of physics at Rochester, in a release. “We’re making a two-terminal memristor device, which drives one type of crystal to another type of crystal phase. Those two crystal phases have different resistance that you can then store as memory.”

By straining the 2D materials, they can be at a point between two different crystal phases and can be nudged in either direction with relatively little power.

“We engineered it by essentially just stretching the material in one direction and compressing it in another,” continued Wu. “By doing that, you enhance the performance by orders of magnitude. I see a path where this could end up in home computers as a form of memory that’s ultra-fast and ultra-efficient. That could have big implications for computing in general.” [2]

Silver-based memristive device

Researchers from Sahmyook University and Yonsei University propose using a silver-dispersive chalcogenide thin film for resistance-switching in memristive devices.

“Our diffusive Ag-based memristive device in a chalcogenide thin film shows low power consumption and mimics the human brain’s parallel processing. This makes it suitable for implementation in crossbar arrays, and it achieved ~92% recognition rate in the MNIST (Modified National Institute of Standards and Technology) handwritten digit recognition database,” said Min Kyu Yang, a professor at Sahmyook University, in a statement.

The device does not require an electric current to induce chemical change before manufacture or operation and demonstrated both state retention and reliable endurance in an environment of 85°C for 2 hours. [3]

References

[1] Bonnet, D., Hirtzlin, T., Majumdar, A. et al. Bringing uncertainty quantification to the extreme-edge with memristor-based Bayesian neural networks. Nat Commun 14, 7530 (2023). https://doi.org/10.1038/s41467-023-43317-9

[2] Hou, W., Azizimanesh, A., Dey, A. et al. Strain engineering of vertical molybdenum ditelluride phase-change memristors. Nat Electron (2023). https://doi.org/10.1038/s41928-023-01071-2

[3] Su Yeon Lee, Jin Joo Ryu, Hyun Kyu Seo, Hyunchul Sohn, Gun Hwan Kim, Min Kyu Yang, Ag-dispersive chalcogenide media for readily activated electronic memristor, Applied Surface Science, Volume 644, 2024, 158747, ISSN 0169-4332, https://doi.org/10.1016/j.apsusc.2023.158747