NAND Read Equalizer With Meta-Generated Threshold Experts
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
NAND flash memory systems face challenges in handling noise-induced errors due to tight voltage levels, leading to increased computational demands and performance issues, particularly in mobile devices, as conventional equalization modules require sophisticated machine-learning models with high latency and memory requirements.
Innovation Solution
Implementing a Multiple Shallow Threshold expert Machine Learning Models (MTM) equalizer that uses multiple shallow machine learning models to classify specific weak decision ranges, allowing for improved soft symbol estimations with reduced memory usage and latency, and is adaptable to various NAND flash memory architectures.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If conventional equalization modules use sophisticated machine-learning models to compensate for noise, then noise compensation accuracy is improved, but computational power requirements and latency increase
Solution Approach 1:
The patent segments the equalization task by dividing the voltage range into multiple weak decision voltage ranges, with each threshold expert model specializing in one specific range. This segmentation allows each model to be simpler and more efficient, reducing overall computational requirements while maintaining accuracy for each segment.
Solution Approach 2:
Each threshold expert model is designed with local quality by specializing in a specific weak decision voltage range rather than attempting to handle all ranges uniformly. This localized approach optimizes performance for each specific range while reducing the complexity burden of handling the entire voltage spectrum.
2Reliability
If sophisticated machine-learning models are used for equalization, then de-noising performance is improved, but memory requirements increase
Solution Approach 1:
By segmenting the equalization task across multiple specialized threshold expert models, each handling a specific voltage range, the patent reduces the memory footprint of individual models compared to a single comprehensive model, while collectively achieving superior de-noising performance.
Solution Approach 2:
The threshold expert models are designed to be computationally lightweight and memory-efficient, sacrificing the complexity of sophisticated deep learning models in exchange for reduced memory requirements and faster execution, suitable for mobile device constraints.
3Quantity of substance
If multiple bits are stored per memory cell to improve manufacturing costs and performance, then storage density is improved, but dynamic voltage range decreases making cells more susceptible to noise
Solution Approach 1:
The patent applies local quality by creating specialized threshold expert models for different weak decision voltage ranges that are affected by noise. Each model is optimized for its specific voltage range, providing targeted noise compensation that preserves the high storage density of multi-bit cells while mitigating their increased noise susceptibility.
Solution Approach 2:
The patent converts the harmful effect of noise in multi-bit cells into a benefit by using the noise characteristics to train specialized threshold expert models. These models learn to compensate for the specific noise patterns affecting each voltage range, turning the previously harmful noise susceptibility into an opportunity for optimized error correction.
4Measurement precision
If threshold expert models are trained for each specific weak decision voltage range, then classification accuracy is improved, but model complexity increases
Solution Approach 1:
The patent segments the classification task across multiple threshold expert models, each handling a specific weak decision voltage range. This segmentation improves classification accuracy for each range while keeping individual model complexity low, as each model only needs to handle a narrow voltage range rather than the entire spectrum.
Solution Approach 2:
Each threshold expert model performs partial action by focusing only on its specific voltage range rather than attempting to classify all possible voltage levels. This partial specialization achieves high accuracy for the relevant range while avoiding the excessive complexity of a universal model.
Data Source
AI summary
Systems and methods of the present disclosure may be used to improve equalization module architectures for NAND cell read information. For example, embodiments of the present disclosure may provide for de-noising of NAND cell read information using a Multiple Shallow Threshold expert Machine Learning Models (MTM) equalizer. An MTM equalizer may include multiple shallow machine learning models. A meta network may generate parameters for each of the shallow machine learning models such that each shallow machine learning model may be able to solve a classification task (e.g., a binary classification task) corresponding to a weak decision range between two possible read information values for a given NAND cell read operation. Accordingly, during inference, each read sample with a read value within a weak decision range may be passed through a corresponding shallow machine learning model (e.g., a corresponding threshold expert) that is associated with the particular weak decision range.