Deep Transfer Learning-based Detection for Flash Memory Channels

TL;DR

This paper introduces a deep transfer learning approach for flash memory data detection that significantly reduces training data requirements and enables unsupervised detection, improving efficiency and applicability.

Contribution

It proposes a model-based deep transfer learning algorithm and an unsupervised domain adaptation method for flash memory detection, addressing data scarcity and label unavailability issues.

Findings

Reduces training samples from 10^6 to less than 10^4.

Achieves near-optimal BER with minimal or no labeled data.

Demonstrates effectiveness through channel error analysis and simulations.

Abstract

The NAND flash memory channel is corrupted by different types of noises, such as the data retention noise and the wear-out noise, which lead to unknown channel offset and make the flash memory channel non-stationary. In the literature, machine learning-based methods have been proposed for data detection for flash memory channels. However, these methods require a large number of training samples and labels to achieve a satisfactory performance, which is costly. Furthermore, with a large unknown channel offset, it may be impossible to obtain enough correct labels. In this paper, we reformulate the data detection for the flash memory channel as a transfer learning (TL) problem. We then propose a model-based deep TL (DTL) algorithm for flash memory channel detection. It can effectively reduce the training data size from $10^6$ samples to less than 104 samples. Moreover, we propose an unsupervised domain adaptation (UDA)-based DTL algorithm using moment alignment, which can detect data without any labels. Hence, it is suitable for scenarios where the decoding of error-correcting code fails and no labels can be obtained. Finally, a UDA-based threshold detector is proposed to eliminate the need for a neural network. Both the channel raw error rate analysis and simulation results demonstrate that the proposed DTL-based detection schemes can achieve near-optimal bit error rate (BER) performance with much less training data and/or without using any labels.