Paper Title: Layer-by-layer Adaptively Optimized ECC for NAND Flash SSD Storing CNN Weights

Paper Abstract: Scene recognition with convolutional neural network (CNN) is a key technology in IoT systems such as self-driving cars [1] The trained weight data are stored in triple-level cell (TLC) NAND flash-based SSD. Error correcting code (ECC) which requires large overhead is essential for TLC NAND flash memory. To save the overhead of ECC [2], this paper proposes Layer-by-layer Adaptively Optimized ECC (LBL-ECC) for trained CNN weights [3]. LBL-ECC is composed of Layer-by-layer Iteration-Optimized Low Density Parity-Check (LBL-LDPC) and Layer-by-layer Code-length Adjusted Asymmetric Coding (LBL-AC). These techniques utilize for the feature of CNN. The 1st proposal, LBL-LDPC is adaptively changing the strength of LDPC ECC according to the importance of CNN weight data. LBL-LDPC can reduce the ECC decoding time. The 2nd proposal, LBL-AC is adaptively optimizing Code Length (CL) Asymmetric Coding (AC) [4, 5] for each convolutional layer. AC improves the reliability of TLC NAND flash memory for modulating the VTH-distribution. As a result, LBL-ECC decreases LDPC decoding time by 14%. Furthermore, LBL-ECC extends the data-retention time by 230%.

Paper Author: Keita Mizushina, Student, Chuo University

Author Bio: Keita Mizushina received the B.E. degree in 2018 from the Department of Electrical, Electronic, and Communication Engineering, Chuo University Tokyo, Japan, where he is currently working toward the M.S. degree. His research is on SSD reliability enhancement, using deep learning technologies. He proposed adaptively optimizing ECC techniques to improve the reliability of TLC NAND flash-based SSDs for scene recognition in IoT edge devices using convolutional neural networks (CNN).