Neural Network Group Integrity Checks for Autonomous Control
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional neural networks face accuracy loss and catastrophic memory issues when operating in varying environments, leading to suboptimal performance as they are trained with increasingly large datasets to cover multiple settings.
Innovation Solution
A system that collects training data for specific settings, associates characteristics with neural network coefficients, and stores this information in a database management system, allowing for adaptive reconfiguration of neural networks based on changing settings by retrieving and instantiating the appropriate coefficients and structures.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Adaptability or versatility
If neural networks are trained with increasingly large datasets to cover multiple settings, then the coverage of different environments is improved, but accuracy loss and catastrophic memory issues occur
Solution Approach 1:
The patent segments the training process by creating multiple specialized neural networks, each trained on data from a specific setting or environment. Instead of one large network attempting to cover all settings, the system divides the problem into smaller segments (different neural networks for different settings), which are then selected based on the current operating context. This segmentation prevents accuracy loss by ensuring each network specializes in its designated domain.
Solution Approach 2:
The patent applies local quality by assigning different characteristics and training data to different neural networks based on their intended operating settings. Each neural network is optimized with local quality tailored to its specific environment rather than using a generic one-size-fits-all approach. This ensures that each network maintains high accuracy for its specific domain while the system as a whole achieves broad adaptability.
2Adaptability or versatility
If neural networks are trained with increasingly large datasets to cover multiple settings, then the coverage of different environments is improved, but catastrophic memory loss occurs
Solution Approach 1:
The patent segments the memory requirements by distributing training data across multiple smaller neural networks rather than loading all data into one large network. Each network stores only the data and parameters necessary for its specific setting, significantly reducing the memory footprint of any single network while maintaining the ability to handle multiple environments through selective instantiation.
Solution Approach 2:
The patent implements discarding and recovering by selectively instantiating and de-instantiating neural networks based on the current operating setting. When a neural network is no longer needed for the current context, it is discarded (de-instantiated), freeing its memory resources. When needed again, it can be recovered (re-instantiated). This dynamic memory management allows the system to handle multiple settings without permanently allocating memory for all possible networks simultaneously.
3Adaptability or versatility
If conventional neural networks operate in varying environments, then adaptability is improved, but performance becomes suboptimal
Solution Approach 1:
The patent applies dynamics by making the neural network configuration adaptive and changeable based on the current operating environment. Instead of a static neural network structure, the system dynamically selects and instantiates the appropriate neural network for the current setting, ensuring that the system always operates with optimally configured parameters for the present conditions. This dynamic adaptation maintains both adaptability and performance efficiency.
Data Source
AI summary
Methods and systems for controlling an autonomous machine. The autonomous machine has sensors generating input data, while the controller includes two or more neural networks that inference using the input data and generate output data. The neural networks can be trained using an identical set of training data set. The output data from each of the neural networks are monitored to ensure that the integrity of the operation is maintained by, for example, the output data from one neural network is compared with the output data from another neural network to verify the consistency. If the comparison yields that the integrity of the system is not maintained at an acceptable level, the controller can stop using the output in controlling the autonomous machine.