Neuromorphic Controller Dynamic Discretization Step Size
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
The use of neuromorphic elements in neural networks leads to quantization errors due to discrete resistance changes, resulting in deteriorated identification accuracy and longer convergence times for weight updates, particularly in applications like handwritten digit recognition.
Innovation Solution
A controller that adjusts the discretization step size of neuromorphic elements based on a true weight value with higher accuracy than the element's resolution, using a dynamic range and offset quantity to minimize errors and improve accuracy, allowing for improved product-sum operations in neural networks.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Power
If a neuromorphic element is used for weight storage and updating in a neural network, then low power consumption and high speed processing are achieved, but identification accuracy deteriorates due to quantization error from discrete resistance changes
Solution Approach 1:
The patent changes the parameter of discretization step size dynamically during the learning process. Initially, a smaller step size is used to achieve finer weight updates and higher accuracy. As learning progresses, the step size is increased to accelerate convergence. This parameter adaptation resolves the contradiction by allowing high precision when needed while maintaining the inherent low power consumption of neuromorphic elements.
Solution Approach 2:
The patent introduces dynamic adjustment of the discretization step size based on learning progress and error metrics. The step size is not fixed but evolves during training, becoming coarser as the network converges. This dynamic approach allows the system to achieve both high initial accuracy and fast final convergence, resolving the trade-off between precision and speed.
2Measurement precision
If a fixed small discretization step size is used for high precision weight updates, then identification accuracy improves, but the period of time required for weight update convergence increases
Solution Approach 1:
The patent makes the discretization step size dynamic rather than fixed. During early learning stages when large weight changes are needed, a larger step size is used. As the network approaches convergence and finer adjustments are needed, the step size is reduced. This dynamic adjustment simultaneously reduces convergence time while maintaining final accuracy.
Solution Approach 2:
The patent implements periodic adjustment of the discretization step size based on learning progress metrics such as error reduction rates. The step size is recalibrated at intervals during training, allowing the system to adapt between aggressive updates (when error is high) and fine-tuned updates (when error is low), optimizing both speed and accuracy throughout the learning process.
Data Source
AI summary
A controller is a controller of an array including a neuromorphic element that multiplies a weight based on a value of a variable characteristic by a signal, and includes a control unit that controls the characteristic of the neuromorphic element by using a discretization step size obtained so that a predetermined condition for reducing an error or a predetermined condition for improving accuracy is satisfied on the basis of a case where a true value of the weight obtained with a higher accuracy than a resolution of the characteristic of the neuromorphic element is used and a case where a discretization step size which is set for the characteristic of the neuromorphic element is used.