Neural Network Weight Quantization for Mobile Object Recognition
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Neural networks for image recognition on mobile devices face challenges due to limited memory and processing resources, which restrict their performance and efficiency in object recognition tasks.
Innovation Solution
A system where a server system trains a neural network using substantial resources and exports the configuration to a client device with limited resources, employing quantization techniques to reduce weight precision and enable efficient object recognition, allowing for real-time image recognition on mobile devices.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If full-precision neural network weights are used on mobile devices, then object recognition accuracy is improved, but memory consumption and processing requirements increase beyond available resources
Solution Approach 1:
The patent applies quantization to change the precision parameter of neural network weights from high precision (e.g., 32-bit floating point) to low precision (e.g., 8-bit integers or lower). This parameter transformation reduces memory consumption and computational requirements while maintaining acceptable recognition accuracy through techniques like fixed-point arithmetic and quantization-aware training
Solution Approach 2:
The patent uses simplified, lower-precision weight representations that require less memory and processing power. These quantized weights act as simplified substitutes for full-precision weights, enabling deployment on resource-constrained mobile devices without requiring the full computational capabilities of server systems
2Measurement precision
If full-precision neural network weights are used on mobile devices, then object recognition accuracy is improved, but power consumption increases beyond acceptable limits
Solution Approach 1:
By changing the numerical precision parameter of weights from high-precision floating-point to low-precision fixed-point or integer representations, the patent reduces the computational energy required for matrix multiplications and neural network inference, directly lowering power consumption on mobile devices
3Measurement precision
If full-precision neural network weights are used on mobile devices, then object recognition accuracy is improved, but device cost increases
Solution Approach 1:
The patent reduces the bit-width parameter of weight storage from 32 bits to 8 bits or lower, which decreases memory capacity requirements and allows the use of cheaper, lower-density memory technologies in mobile device manufacturing, thereby reducing overall device cost
4Measurement precision
If training data set size is increased to improve neural network performance, then object recognition accuracy is improved, but memory and processing requirements during training increase beyond mobile device capabilities
Solution Approach 1:
The patent divides the training process into two segments: (1) initial training with large datasets on server systems with abundant resources, and (2) subsequent quantization and fine-tuning on mobile devices. This segmentation allows mobile devices to benefit from large-scale training without requiring them to handle the computational burden of training on massive datasets
Solution Approach 2:
The patent performs preliminary training and quantization on server systems before deploying the model to mobile devices. This preliminary action prepares the neural network weights in advance, allowing mobile devices to skip the resource-intensive training phase and directly use the pre-quantized model for inference
Data Source
AI summary
A client device configured with a neural network includes a processor, a memory, a user interface, a communications interface, a power supply and an input device, wherein the memory includes a trained neural network received from a server system that has trained and configured the neural network for the client device. A server system and a method of training a neural network are disclosed.