Scalable Neural Network Training for Edge Device Inference Accuracy
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Neural architecture search techniques lack guidelines for varying training processes, leading to insufficient inference accuracy due to the absence of specific considerations for subject devices, such as edge devices, which require tailored data resolutions and network structures.
Innovation Solution
A learning apparatus that determines multiple data resolutions and corresponding neural network structures based on device information, including data resolution and processing capabilities, to train scalable networks that can adapt to different input data resolutions, thereby generating a trained model optimized for specific subject devices.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Device complexity
If neural architecture search is used to optimize network architecture, then device complexity is reduced through automated design, but inference accuracy deteriorates due to lack of subject device-specific considerations
Solution Approach 1:
The patent applies parameter changes by systematically varying training data resolutions around the subject device's native resolution (e.g., if subject device uses 640x640, training uses 512x512, 640x640, 768x768). This resolves the contradiction by making the training process adaptive to device characteristics while maintaining automated NAS, thereby improving inference accuracy without increasing design complexity.
Solution Approach 2:
The patent performs preliminary action by determining device information (resolution, processing capabilities) before the training process begins. This pre-analysis allows the training configuration to be optimized for the specific subject device, resolving the accuracy issue while keeping the NAS automation intact.
2Ease of operation
If training is performed with fixed data resolution, then training process simplicity is maintained, but adaptability deteriorates due to inability to accommodate different device specifications
Solution Approach 1:
The patent applies dynamics by making the training data resolution variable rather than fixed. The system dynamically selects multiple resolutions based on the subject device's characteristics, allowing the training process to adapt to different devices while maintaining a standardized automated workflow. This resolves the contradiction between simplicity and adaptability.
Solution Approach 2:
The patent achieves universality by creating a training framework that works across multiple device types and resolutions. The same automated NAS process can be applied to any subject device by simply changing the device information input, making the system both simple to operate and highly adaptable.
3Loss of time
If subject device specifications are not considered during training, then training time is reduced, but inference accuracy deteriorates due to mismatch between training and deployment conditions
Solution Approach 1:
The patent changes the training parameter (data resolution) to match the subject device's specifications. By training with resolutions that cover the device's native resolution (e.g., including 640x640 if that's the device resolution), the model achieves better accuracy without excessive time penalty, as the resolutions are strategically selected rather than exhaustively tested.
Solution Approach 2:
The patent applies partial action by selecting a limited set of resolutions that are most relevant to the subject device (centered around the native resolution) rather than training with all possible resolutions. This provides sufficient accuracy improvement without the full time cost of exhaustive resolution testing.
Data Source
AI summary
According to one embodiment, a learning apparatus includes a processor. The processor determines, based on a data resolution of subject data obtained at a subject device, a plurality of data resolutions that differ from one another within a range covering the data resolution of the subject data, the data resolutions each indicating a corresponding amount of information per unit. The processor trains a scalable network with training samples corresponding to each of the plurality of data resolutions, the scalable network being a neural network adapted to change a data resolution of input data.