Neural Network Structure Search for Hardware-Constrained Edge Devices
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional techniques struggle to design neural network models that are optimized for specific hardware, such as mobile and in-vehicle devices, due to high memory usage and calculation demands, making real-time processing difficult.
Innovation Solution
A learning device with a structure search unit, parameter search unit, pruning unit, and morphing unit that searches for and optimizes neural network structures and parameters based on target hardware constraints, such as model size, calculation amount, and power consumption, using Bayesian search methods to select suitable search space information and hyperparameters.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If conventional neural network models with deep learning are used, then recognition performance is improved, but memory usage and calculation amount increase making real-time processing difficult
Solution Approach 1:
The patent extracts and removes unnecessary components from the neural network model through pruning operations. Specifically, it identifies and removes redundant neurons, connections, and layers that do not contribute significantly to recognition performance, thereby reducing memory usage while preserving essential functional capabilities.
Solution Approach 2:
The patent employs approximate computation techniques that use simpler, less resource-intensive calculations in place of expensive precise computations. This includes using quantized weight representations and approximate inference methods that trade off minimal precision for significant reductions in memory requirements and computational burden.
2Measurement precision
If conventional neural network models with deep learning are used, then recognition performance is improved, but calculation amount increases making real-time processing difficult
Solution Approach 1:
The patent segments the neural network into distinct functional modules and processing stages. By dividing the network into separate convolutional blocks, pooling layers, and fully connected layers, it enables selective optimization and efficient resource allocation, reducing overall computational burden while maintaining recognition accuracy.
Solution Approach 2:
The patent applies partial computation strategies where not all neurons and connections are activated during inference. Through techniques such as channel pruning and neuron dropout, it disables portions of the network that contribute minimally to the final output, thereby reducing calculation amount while preserving sufficient recognition performance.
3Quantity of substance
If model downsizing methods such as deleting units are applied, then model size is reduced, but optimization in accordance with target hardware becomes difficult
Solution Approach 1:
The patent implements dynamic model adaptation that adjusts network architecture based on target hardware characteristics. Through automated hyperparameter search and pruning strategies tailored to specific device constraints (memory capacity, processing speed, power availability), the model dynamically optimizes its structure to achieve optimal performance on the intended deployment platform.
Solution Approach 2:
The patent systematically varies key model parameters such as layer depth, filter counts, kernel sizes, and activation functions to find configurations that balance model size with hardware compatibility. By searching through parameter spaces and selecting optimal values, it achieves both downsizing and hardware-specific optimization simultaneously.
Data Source
AI summary
A learning device includes a structure search unit that searches for a first learned model structure obtained by selecting search space information in accordance with a target constraint condition of target hardware for each of a plurality of convolution processing blocks included in a base model structure in a neural network model; a parameter search unit that searches for a learning parameter of the neural network model in accordance with the target constraint condition; and a pruning unit that deletes a unit of at least one of the plurality of convolution processing blocks in the first learned model structure based on the target constraint condition and generates a second learned model structure.