Memristor Write Signal Control for Trapping-Resistant Weight Updates
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Memristors in neuromorphic devices can experience a trapping phenomenon where the magnetic domain wall does not move as expected due to manufacturing limitations, leading to undesired conductance changes during learning processes, hindering the optimization of neural networks.
Innovation Solution
A control unit determines a writing signal based on the necessary conductance change amount calculated from the update value of the neural network weight and an expected value, considering factors like memristor position and temperature, to ensure appropriate conductance change even when trapping phenomena occur.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If a writing signal is applied to a magnetic domain wall moving element to change conductance, then the conductance should change according to the weight update value of the neural network, but the magnetic domain wall may be trapped by potential changes due to manufacturing grooves, causing the conductance not to change as desired
Solution Approach 1:
The control unit calculates the necessary writing signal strength in advance based on the conductance change amount required for the neural network weight update. By determining the appropriate writing signal parameters before application, the system prepares to overcome potential trapping barriers that may exist due to manufacturing grooves, ensuring the magnetic domain wall can be moved successfully to achieve the desired conductance change.
Solution Approach 2:
The system uses feedback by calculating the writing signal based on the relationship between the applied writing signal and the resulting conductance change. The control unit adjusts the writing signal strength according to the necessary conductance change amount, creating a feedback mechanism that accounts for the stochastic trapping phenomenon and ensures reliable weight updates in the neural network.
2Reliability
If the writing signal strength is increased to overcome the trapping phenomenon, then the conductance change reliability improves, but the energy consumption and risk of damaging the memristor increases
Solution Approach 1:
The control unit dynamically adjusts the writing signal parameters (strength, duration) based on the necessary conductance change amount calculated from the neural network weight update requirements. By optimizing these parameters rather than using a fixed high-strength signal, the system achieves reliable conductance changes while minimizing energy consumption and avoiding memristor damage.
3Productivity
If the writing signal is applied frequently to update neural network weights during learning, then the learning speed improves, but the trapping phenomenon occurs more often, reducing the effectiveness of weight updates
Solution Approach 1:
The control unit implements a feedback mechanism that calculates the writing signal strength based on the necessary conductance change amount for each weight update. This feedback approach ensures that even with frequent updates during learning, each writing signal is optimized to overcome potential trapping effects and achieve the desired conductance change, maintaining both learning speed and update effectiveness.
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
The solution enables effective conductance changes in memristors, ensuring optimal neural network learning and performance by addressing the stochastic conductance issues caused by trapping phenomena.
Implementation Method 1
The magnetic domain wall can be moved by applying a writing signal (a current) to the magnetic domain wall moving layer. A resistance value in a stacking direction of the magnetic domain wall moving element changes according to the position of the magnetic domain wall.
Implementation Method 2
The tunnel insulating layer is interposed between the magnetic domain wall moving layer and the magnetization fixed layer.
Data Source
AI summary
A neuromorphic device includes a control unit. The control unit is configured to be connectable to a memristor having electrical characteristics in which conductance change occurs stochastically when a writing signal is applied to the memristor and is configured to apply the writing signal to the memristor. The writing signal is determined based on a necessary value of the conductance change amount of the memristor which is calculated from an update value of a weight of a neural network and an expected value by which the conductance of the memristor changes when a reference writing signal is applied to the memristor.