Neural Network Construction Device for Balancing Inference Time and Accuracy
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing neural network construction devices struggle to balance inference speed and accuracy while operating with limited computational resources, often sacrificing either necessary time or accuracy.
Innovation Solution
A neural network construction device that allocates operations to layers based on construction conditions, generates operation combinations, trains the model, and adjusts weights to meet both requested inference time and accuracy by utilizing processor and memory to execute programs that specify and determine optimal operation allocations.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If the operation amount of the learning model is reduced to satisfy computational resource constraints, then the model can be executed on devices with limited resources, but the necessary time for inference may not be satisfied
Solution Approach 1:
The patent applies dynamics by making the neural network structure adjustable and reconfigurable. Different operation combinations are generated and selected based on the specific balance between computational resources and inference time requirements. The system dynamically chooses the optimal model configuration rather than using a fixed structure, allowing adaptation to different deployment scenarios with varying resource constraints and performance requirements.
Solution Approach 2:
The patent changes structural parameters of the neural network by generating multiple operation combinations with different architectures. Each combination represents a different configuration of operations across layers, effectively changing the model's parameters to achieve the desired trade-off between computational efficiency and inference speed. This parameter exploration allows finding configurations that satisfy both resource constraints and time requirements.
2Productivity
If the operation amount is reduced to meet resource constraints, then the model becomes more deployable on resource-limited devices, but the accuracy of recognition may deteriorate
Solution Approach 1:
The patent systematically explores different structural parameters of the neural network through multiple operation combinations. By adjusting the configuration of operations in each layer, the system finds optimal parameter settings that maintain recognition accuracy while reducing computational complexity. This parameter optimization ensures that the simplified model does not sacrifice too much accuracy for the sake of efficiency.
Solution Approach 2:
The patent incorporates a determination process that evaluates whether the generated model satisfies both the operation amount constraints and the accuracy requirements. This feedback mechanism allows the system to assess the performance of each operation combination and select configurations that meet the dual criteria, ensuring that accuracy deterioration is prevented while achieving computational efficiency.
3Ease of manufacture
If a fixed neural network structure is used, then the model construction is simple, but it cannot simultaneously satisfy both inference time and accuracy requirements under resource constraints
Solution Approach 1:
The patent segments the neural network into multiple layers with different operable operations that can be independently configured. This segmentation allows the system to optimize each layer's operations separately to achieve the global optimization of the entire network. By dividing the model construction into manageable segments, the system maintains relative simplicity while enabling sophisticated performance tuning through multiple operation combinations.
Data Source
AI summary
When a first determination unit determines that a total necessary time does not satisfy a requested necessary time, or when a second determination unit determines that accuracy of inference does not satisfy requested inference accuracy, a generation unit generates a second operation combination different from a first operation combination.