Neuromorphic Product-Sum Device Bias Element Segmentation
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing neural networks face performance degradation when a bias term element malfunctions, as the value of 'weight' is set for specific combinations, whereas the bias term biases all values of the layer, leading to significant performance reduction.
Innovation Solution
A product-sum operation device with variable-input and fixed-input product operation elements, where the fixed-input elements can disconnect upon output current increase malfunctions, and a malfunction diagnosis unit resets resistance values to maintain network performance.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If bias term elements are used in neural networks, then the network can perform product-sum operations effectively, but the performance degrades significantly when a bias term element malfunctions
Solution Approach 1:
The bias term elements are segmented into multiple independent elements (first bias term element and second bias term element) rather than using a single element. This segmentation ensures that if one element malfunctions, the other can still provide the bias term function, thereby reducing the impact of individual element failures on overall network performance.
2Reliability
If multiple bias term elements are used to improve reliability, then the network performance is maintained during malfunctions, but the device complexity increases
Solution Approach 1:
The bias term is divided into multiple independent elements, with at least one being a first bias term element and another being a second bias term element. This segmentation provides redundancy without requiring a complete redesign of the bias term mechanism, achieving reliability improvement with minimal additional complexity.
3Device complexity
If fixed-input product operation elements are used for bias terms, then the structure is simplified, but the elements cannot be reset after malfunction
Solution Approach 1:
The bias term elements are designed with dynamic reset capability through the relearning unit, which can reset the resistance values of the bias term elements after malfunction. This transforms the static fixed-input product operation elements into dynamic components that can be restored to their initial states, enabling repair and maintenance.
Solution Approach 2:
The relearning unit provides feedback control by detecting the resistance values of the bias term elements and resetting them when malfunction is detected. This feedback mechanism enables the system to automatically recover from failures, improving ease of repair while maintaining the simplified fixed-input structure.
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 effectively curbs performance reduction in neural networks by diagnosing and addressing malfunctioning bias term elements, ensuring continued discrimination performance through relearning processes.
Implementation Method 1
each of the plurality of variable-input product operation elements and the plurality of fixed-input product operation elements is a resistance change element
Implementation Method 2
the resistance change element may be a magnetoresistance effect element exhibiting a magnetoresistance effect
Data Source
AI summary
A product-sum operation device includes a product operator and a sum operator. The product operator includes a plurality of variable-input product operation elements and a plurality of fixed-input product operation elements. Each of the plurality of variable-input product operation elements and the plurality of fixed-input product operation elements and is a resistance change element. The product-sum operation device includes variable input units and that input a variable signal to a plurality of variable-input product operation elements and fixed input units and that input a determined signal to the plurality of fixed-input product operation elements and in synchronization with the variable signal. The sum operator includes an output detector that determines the sum of outputs from the plurality of variable-input product operation elements and outputs from the plurality of fixed-input product operation elements.