Non-idealities in artificial synapses | Nature Reviews Physics

Source Domain: www.nature.com

The article discusses various types of memory devices proposed for neuromorphic computing and their application in deep learning.
Key types of memory devices mentioned include resistive random-access memory (RRAM), phase-change memory (PCM), magnetic tunnel junctions (MTJs), and ferroelectric field-effect transistors (FeFET).
The work emphasizes the importance of endurance, retention, and linearity of these devices for effective neuromorphic and deep learning applications.
Recent advances in materials science, such as incorporating HfOx, Ge2Sb2Te5, and other nanocomposites, are highlighted for their beneficial properties for neuromorphic devices.
The discussion extends to challenges such as process variation, linearity, symmetry, and resistance drift, and how these affect the practical implementation and performance of neuromorphic systems.
The article also touches on the potential applications of these devices in brain-inspired computing and unconventional computing paradigms.
In-memory computing and the integration of memory and processing within a single device architecture are highlighted as key strategies for enhancing the efficiency of deep learning computations.
The potential of probabilistic and stochastic behavior in devices like memristors for enabling Bayesian inference and machine learning tasks is explored.

You may have missed