Neural Network Generation Device for Embedded IoT Devices
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Convolutional neural networks (CNNs) used in image recognition and similar applications face challenges in efficiently running on embedded devices like IoT devices, requiring optimized circuit and model generation for high-speed and efficient operations.
Innovation Solution
A neural network generation device and method that generates execution models and software for CNN operations based on hardware and network information, partitioning operations to optimize performance and efficiency on embedded devices.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Adaptability or versatility
If convolutional neural networks are used in embedded devices, then image recognition capability is improved, but device complexity increases
Solution Approach 1:
The patent segments the neural network into multiple layers (convolution layers, pooling layers, fully connected layers) and processes data through each layer sequentially. This segmentation allows the complex image recognition task to be broken down into manageable operations that can be executed on embedded devices with limited resources, thus resolving the contradiction between capability and complexity.
Solution Approach 2:
The patent implements dynamic control of the neural network execution by adjusting operation sequences and resource allocation based on hardware capabilities. The system dynamically manages memory usage, processing order, and data flow to optimize performance on embedded devices, balancing the image recognition capability with device complexity constraints.
2Measurement precision
If multilayered neural network structure is implemented, then recognition accuracy is improved, but operation speed decreases
Solution Approach 1:
The patent performs preliminary processing of input images including normalization, scaling, and feature extraction before feeding data into the multilayered neural network. This preliminary action reduces the computational burden on subsequent layers, maintaining recognition accuracy while improving operation speed on embedded devices.
Solution Approach 2:
The patent implements continuous data flow through the neural network layers with optimized memory management and caching mechanisms. By keeping data in accessible memory states and minimizing I/O operations between layers, the system maintains high recognition accuracy while improving operation speed through continuous processing.
3Adaptability or versatility
If many convolution operations are performed, then feature extraction capability is improved, but energy consumption increases
Solution Approach 1:
The patent extracts only the most essential features at each layer rather than performing exhaustive convolution operations on all data. By selectively extracting salient features and discarding redundant information, the system maintains strong feature extraction capability while significantly reducing energy consumption on battery-powered embedded devices.
Solution Approach 2:
The patent implements partial convolution operations where not all filter kernels are applied to all input regions, especially in later layers. This partial action approach maintains sufficient feature extraction capability for accurate recognition while reducing the total number of operations and associated energy consumption.
Data Source
AI summary
A neural network generation device that generates a neural network execution model for performing neural network operations, the neural network generation device including an execution model generation unit that generates the neural network execution model based on hardware information regarding hardware in which the neural network execution model is running and network information regarding the neural network, and a software generation unit that generates software for running neural network hardware obtained by installing the neural network model in the hardware.