Neural Network Generation Device for Embedded IoT Data Quantization
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Convolutional neural networks (CNNs) used in embedded devices, such as IoT devices, face challenges in efficiently operating and generating circuits and models that are optimized for hardware configurations, requiring high-speed and efficient operation methods and software generation programs.
Innovation Solution
A neural network generation device that generates a neural network execution model capable of converting input data with S bits or more into fewer bits by comparing with multiple threshold values, optimizing operations for embedded devices.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Adaptability or versatility
If convolutional neural networks are used in embedded devices, then image recognition functionality is achieved, but computational resources and power consumption increase
Solution Approach 1:
The patent applies parameter changes by converting input data from high-bit precision (S bits or more) to low-bit precision (fewer bits than S) through quantization. This changes the data representation parameters to reduce computational complexity and power consumption while maintaining neural network functionality in embedded devices
Solution Approach 2:
The patent segments the data conversion process into multiple stages using multiple threshold values. The quantization is performed in segments by comparing input data with multiple threshold values to produce converted values with fewer bits, making the computation more efficient for embedded devices
2Measurement precision
If high-bit precision data is processed, then measurement precision is maintained, but data redundancy and processing complexity increase
Solution Approach 1:
The patent changes the bit-width parameter of data representation from high-bit (S bits or more) to low-bit (fewer bits) through quantization. This parameter change reduces data redundancy and processing complexity while maintaining sufficient precision for neural network operations in embedded devices
Solution Approach 2:
The patent uses a simplified quantization approach that replaces complex high-bit data processing with simpler low-bit operations. The quantized data with fewer bits serves as a disposable representation that reduces computational burden while maintaining functional accuracy
3Ease of manufacture
If circuits and models are generated for embedded devices, then hardware optimization is achieved, but generation complexity increases
Solution Approach 1:
The patent performs preliminary quantization of input data before it enters the neural network processing pipeline. By pre-converting high-bit data to low-bit data using multiple threshold values, the system prepares optimized data representations in advance, simplifying subsequent hardware implementation and reducing generation complexity
Data Source
AI summary
A neural network generation device that generates a neural network execution model for performing operations of a neural network wherein the neural network execution model converts input data including elements with 8 bits or more to converted values with fewer bits than the elements, based on comparisons with multiple threshold values.