What is it about?
ReRAMs can be used to accelerate deep learning in future. However, there are reliability issues. Common solutions for reliability are often too expensive. In this work, we show that regularization methods, that are native to deep learning, can be repurposed to improve hardware reliability.
Featured Image
Photo by Christian Wiediger on Unsplash
Why is it important?
The idea of repurposing regularization for fault-tolerance on ReRAMs is unique here. Prior reliability solutions often introduce additional operations and are costly to implement. The use of regularization can solve this problem. Since regularization methods are native to deep learning, their use does not introduce any additional operations. We show that ReRAMs can be made reliable by using these existing mechanisms at significantly lower cost.
Perspectives
Deep learning practitioners use regularizers to make their models more robust. Here, we have shown that we can repurpose it for solving hardware reliability challenges as well. The use of an already existing mechanism leads to low cost and a simple solution. I believe that this makes it a very attractive solution for future computing systems and should be adopted in commercial products.
Biresh Kumar Joardar
University of Houston
Read the Original
This page is a summary of: Fault-Tolerant Deep Learning Using Regularization, October 2022, ACM (Association for Computing Machinery),
DOI: 10.1145/3508352.3561120.
You can read the full text:
Contributors
The following have contributed to this page
