Analog Memory Architecture for Zero-Value Shifting
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current memory architectures for neural networks face challenges in efficiently implementing zero-value shifting and differential reading, which are crucial for neural network training processes like backpropagation, especially in achieving symmetry point convergence and handling positive and non-positive weights.
Innovation Solution
An analog memory architecture with a weight array and a reference array, utilizing cross-point devices and differential unipolar switching memory devices, enables zero-value shifting by applying voltage pulses and copying conductance values to achieve symmetry point convergence and represent both positive and non-positive weights through double differential reading.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If traditional memory architectures are used for neural network training, then computational tasks can be performed, but efficiency is insufficient and zero-value shifting cannot be achieved
Solution Approach 1:
The memory architecture is segmented into a weight array for storing neural network weights and a reference array for storing reference values. This segmentation enables independent optimization of each array's function, allowing efficient zero-value shifting through differential reading while maintaining manageable complexity through modular design
Solution Approach 2:
A differential reading mechanism is introduced as an intermediary between the weight array and readout circuitry. This intermediary performs differential comparisons that enable zero-value shifting capability without requiring complex modifications to the base memory cells, thus improving productivity while controlling complexity
2Measurement precision
If symmetry point convergence is implemented, then weight representation accuracy is improved, but additional computational steps are required
Solution Approach 1:
The architecture pre-establishes symmetry points during initialization and maintains them throughout training operations. By preparing the reference array with pre-computed reference values corresponding to symmetry points, the system achieves accurate weight representation without requiring iterative convergence steps during forward propagation, thus improving precision without time penalty
Solution Approach 2:
The reference array stores copied reference values that represent symmetry points. These pre-copied reference values enable direct comparison with weight array values to achieve accurate weight representation, eliminating the need for repeated convergence computations and reducing time loss
Applied Scientific Principles
This section explains which scientific principles are used to turn an abstract innovation direction into a practical engineering solution.
Function Achieved in This Case
This architecture accelerates neural network training by enabling efficient zero-value shifting and differential reading, improving performance over traditional CPU and GPU computational methods through parallel computations and optimized weight representation.
Implementation Method 1
The second cross-point devices include differential unipolar switching memory devices configured to enable zero-value shifting of the outputs of the first cross-point devices
Data Source
AI summary
A system includes an analog memory architecture for performing differential reading. The analog memory architecture includes a weight array including first cross-point devices located at intersections of a first set of conductive column wires and a first set of conductive row wires, and a reference array operatively coupled to the weight array and including second cross-point devices located at intersections of a second set of conductive column wires and a second set of conductive row wires. The second cross-point devices include differential unipolar switching memory devices configured to enable zero-value shifting of the outputs of the first cross-point devices.