Reconfigurable Neural Network Using Correlated Electron Switches
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing artificial neural networks (ANNs) formed with filamentary resistive elements, such as memristors and ReRAM, are unsuitable for use in resource-constrained devices like IoT devices due to their analogue nature, which requires in-situ training and re-tuning for different applications, making them inefficient and impractical for cloud-based or remote neural network training.
Innovation Solution
The implementation of a reconfigurable artificial neural network using non-filamentary non-volatile memory elements, specifically correlated electron switches (CES), which allow for digital encoding of weights and enable ANN operation without in-situ training, allowing for quick and efficient deployment in IoT devices by downloading pre-trained weights from a remote server.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Ease of manufacture
If filamentary resistive elements (memristors, ReRAM) are used to form ANNs, then the neural network can be implemented in hardware, but the analogue nature requires in-situ training and re-tuning for different applications
Solution Approach 1:
The patent changes the fundamental parameter of memory element type from filamentary resistive elements to non-filamentary non-volatile memory elements. This parameter change enables digital encoding of weights instead of analogue resistance values, allowing the same hardware to be reconfigured for different applications by loading pre-trained weight values without requiring in-situ training or re-tuning.
2Reliability
If in-situ training is performed for different applications, then the neural network can be optimized for specific tasks, but the process is expensive and time-consuming
Solution Approach 1:
The patent applies preliminary action by performing the training process beforehand to generate pre-trained weight values, which are then stored and can be quickly loaded into the non-volatile memory elements. This eliminates the need for time-consuming in-situ training when deploying the neural network for specific tasks, as the optimization work has already been completed in advance.
3Duration of action of stationary object
If non-volatile memory elements are used to store weights, then the ANN can operate without continuous power, but the digital encoding requires switching multiple elements to implement weights
Solution Approach 1:
The patent applies segmentation by dividing the weight storage function across multiple non-volatile memory elements at each neural network node. Instead of using a single element, multiple elements work together to represent weight values, enabling digital encoding while maintaining the power retention benefits of non-volatile memory. This segmented approach allows for more flexible and precise weight representation.
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 enables efficient, low-power, and energy-efficient implementation of ANN in IoT devices, providing immunity to resistance variability and allowing for easy reconfiguration across different applications without the need for expensive and time-consuming retraining.
Implementation Method 1
each non-volatile memory element is switchable between a first impedance state and a second impedance state
Data Source
AI summary
Broadly speaking, embodiments of the present techniques provide a reconfigurable hardware-based artificial neural network, wherein weights for each neural network node of the artificial neural network are obtained via training performed external to the neural network.