Analog Crosspoint Array Zero-Weight Alignment
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Resistive devices in analog crosspoint arrays exhibit weight update imbalances due to linear dependence on current weight values when voltage pulses are applied stepwise, leading to asymmetrical adjustments and affecting neural network performance.
Innovation Solution
The technique involves applying repeated voltage pulses to crosspoint devices in the weight array until they converge to their symmetry point, and then copying these conductance values to a reference array to establish a zero-weight point, which balances the device responses and compensates for imbalances.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Ease of operation
If voltage pulses are applied stepwise to adjust weight values in resistive devices, then weight updates can be performed, but the amount of weight update is linearly dependent on the current weight value causing imbalance between up and down adjustments
Solution Approach 1:
The patent applies preliminary action by performing a convergence process before actual computation, where repeated voltage pulses are applied to drive crosspoint devices to their symmetry points. This preliminary convergence establishes a balanced baseline state, ensuring that subsequent weight updates operate from a symmetric reference point, thereby eliminating the linear dependence imbalance.
Solution Approach 2:
The patent utilizes parameter changes by dynamically adjusting the conductance values of crosspoint devices through controlled voltage pulse application. By changing the conductance parameter until convergence at the symmetry point is achieved, the system transforms the unbalanced linear dependence into a balanced relationship, where equal magnitude voltage pulses produce equal magnitude conductance changes in opposite directions.
2Manufacturing precision
If repeated voltage pulses are applied to converge crosspoint devices to symmetry point, then weight update balance is improved, but additional computation time and pulses are required
Solution Approach 1:
The patent applies periodic action through the use of repeated voltage pulses applied in cycles until convergence is achieved. This periodic pulsing continues systematically until all crosspoint devices reach their symmetry points, providing a structured and predictable convergence process that balances precision requirements with time management.
Solution Approach 2:
The patent implements feedback by continuously monitoring the conductance values of crosspoint devices during the voltage pulse application process. This feedback mechanism allows the system to detect when convergence to the symmetry point has been achieved, enabling the process to terminate efficiently without unnecessary additional pulses, thus reducing the time loss while maintaining precision.
3Manufacturing precision
If conductance values are copied from weight array to reference array, then zero-weight point is established, but additional array and copying process are required
Solution Approach 1:
The patent directly applies the copying principle by creating a reference array that replicates the conductance values from the weight array after convergence. This copying process establishes an accurate zero-weight reference point that can be used for subsequent differential measurements, improving precision while managing complexity through systematic duplication of the converged state.
Solution Approach 2:
The reference array serves as an intermediary structure that mediates between the weight array and the computation process. By introducing this intermediate reference array with copied conductance values, the system enables accurate zero-point reference without directly modifying the weight array structure, thus managing complexity while achieving precision.
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 approach improves neural network performance by ensuring balanced and symmetrical weight updates, effectively addressing the imbalance issue and enhancing the accuracy of vector-matrix computations.
Implementation Method 1
applying repeated voltage pulses to the crosspoint devices in the weight array until all of the crosspoint devices in the weight array converge to their own symmetry point
Implementation Method 2
copying conductance values for each crosspoint device from the weight array to the reference array
Data Source
AI summary
Zero-shifting techniques in analog crosspoint arrays are provided. In one aspect, an analog array-based vector-matrix multiplication includes: a weight array connected to a reference array, each including a crossbar array having a set of conductive row wires and a set of conductive column wires intersecting the set of conductive row wires, and optimizable crosspoint devices at intersections of the set of conductive column wires and the set of conductive row wires. A method for analog array-based vector-matrix computing is also provided that includes: applying repeated voltage pulses to the crosspoint devices in the weight array until all of the crosspoint devices in the weight array converge to their own symmetry point; and copying conductance values for each crosspoint device from the weight array to the reference array.