Memristor Crossbar Circuit Temperature Weight Update
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Artificial neural network (ANN) circuits with memristor-based crossbar circuits face performance deterioration due to temperature variations, leading to increased recognition errors in applications like image recognition, as the conductance values of memristors change at different rates with environmental temperature changes, disrupting the relations among weights.
Innovation Solution
Incorporating a temperature sensor to detect environmental temperature changes and an update mechanism that adjusts the trained weights in the crossbar circuit to maintain optimal performance across varying temperatures, either by re-learning weights suitable for specific temperature ranges or updating only the bias signal weights.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If trained weights are fixed in the crossbar circuit, then the circuit structure is simple, but recognition accuracy deteriorates under temperature variations
Solution Approach 1:
The patent implements dynamic weight adjustment by introducing a control circuit that modifies the weight values stored in the crossbar circuit based on detected temperature conditions. Instead of fixed weights, the system dynamically switches between different weight sets (first trained weights and second trained weights) according to temperature ranges, thereby maintaining recognition accuracy across varying thermal environments without requiring a completely reconfigurable circuit architecture.
Solution Approach 2:
The patent changes the parameter of weight values in response to temperature variations. By detecting temperature conditions and switching between different trained weight sets optimized for specific temperature ranges, the system adjusts the critical parameter (weight values) to compensate for temperature-induced performance degradation, thus maintaining reliability without fundamental structural changes.
2Reliability
If multiple trained weights are stored for different temperature ranges, then recognition accuracy is maintained across temperatures, but memory requirements and system complexity increase
Solution Approach 1:
The patent segments the temperature operating range into multiple discrete temperature ranges, each associated with a specific trained weight set. Instead of storing a continuous spectrum of weights, the system divides the temperature domain and assigns optimized weights to each segment, reducing the total memory requirement while maintaining accuracy within each segmented range.
Solution Approach 2:
The patent performs preliminary training to generate multiple weight sets optimized for different temperature ranges before deployment. These pre-computed weight sets are stored in memory, and during operation, the system simply retrieves the appropriate pre-trained weights based on current temperature, avoiding the need for real-time weight computation or extensive online learning resources.
3Reliability
If weight updates are performed frequently to adapt to temperature changes, then performance is maintained, but energy consumption and processing time increase
Solution Approach 1:
The patent implements periodic weight updates based on temperature thresholds rather than continuous adjustment. The system monitors temperature and triggers weight switching only when temperature crosses predefined thresholds between different temperature ranges. This periodic update strategy maintains performance consistency while minimizing the frequency of weight update operations and associated processing time.
Solution Approach 2:
The patent implements dynamic weight adjustment by introducing a control circuit that modifies the weight values stored in the crossbar circuit based on detected temperature conditions. Instead of fixed weights, the system dynamically switches between different weight sets (first trained weights and second trained weights) according to temperature ranges, thereby maintaining recognition accuracy across varying thermal environments without requiring a completely reconfigurable circuit architecture.
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 configuration effectively suppresses the influence of conductance value variations in memristors, reducing recognition errors and maintaining performance across a range of temperatures, from room temperature to high temperatures, by updating weights to match environmental conditions.
Implementation Method 1
multiple memristors that are disposed at respective intersections of the input bars and the output bars to give a weight to the signal to be transmitted, as a variable resistance memory
Implementation Method 2
a temperature sensor to detect environmental temperature
Data Source
AI summary
The present disclosure describes an artificial neural network circuit including: at least one crossbar circuit to transmit a signal between layered neurons of an artificial neural network, the crossbar circuit including multiple input bars, multiple output bars arranged intersecting the input bars, and multiple memristors that are disposed at respective intersections of the input bars and the output bars to give a weight to the signal to be transmitted; a processing circuit to calculate a sum of signals flowing into each of the output bars while a weight to a corresponding signal is given by each of the memristors; a temperature sensor to detect environmental temperature; and an update portion that updates a trained value used in the crossbar circuit and/or the processing circuit.