Dynamic Fixed-Point CNN Model for Embedded Memory Reduction
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Resource-limited embedded systems, such as mobile devices, face challenges in achieving high performance with large-scale convolutional neural networks (CNNs) due to limited memory storage and computing capabilities.
Innovation Solution
A floating-to-fixed arithmetic framework system that transforms a pre-trained CNN model into a dynamic fixed-point model by determining and applying dynamic fixed-point formats for weights, biases, and activations, optimizing the model to fit hardware constraints through a series of operations including summing, truncating, and combining outputs across CNN layers.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If high-precision arithmetic operators are used in CNN architectures, then performance is improved, but device complexity and resource requirements increase
Solution Approach 1:
The patent transforms the precision parameter of arithmetic operators from high-precision floating-point to low-precision fixed-point representation. This parameter change reduces the complexity of arithmetic operators while maintaining acceptable CNN performance through dynamic fixed-point format optimization that adapts to different layers and operations.
Solution Approach 2:
The patent introduces dynamic fixed-point formats where the precision and range parameters are not fixed but adapt dynamically based on the specific CNN layer, operation type, and data distribution. This allows the system to maintain high performance where needed while reducing complexity where possible, resolving the contradiction between performance and device complexity.
2Productivity
If large-scale CNN models are deployed, then performance is improved, but memory storage requirements increase
Solution Approach 1:
The patent changes the data representation parameter from floating-point to fixed-point format for storing weights, biases, and activations in the CNN model. This parameter change significantly reduces the memory storage requirements while maintaining model performance through optimized fixed-point precision that is sufficient for the specific application.
Solution Approach 2:
The patent employs low-precision fixed-point representations that use less memory space, effectively replacing high-precision floating-point data structures with more compact representations. This allows large-scale CNN models to be deployed in resource-constrained environments with limited memory capacity.
3Quantity of substance
If low-precision fixed-point formats are used, then resource requirements are reduced, but computational precision decreases
Solution Approach 1:
The patent implements dynamic fixed-point formats where the precision parameters (integer and fractional bits) are adjusted dynamically based on the specific requirements of each CNN layer and operation. This dynamic adaptation ensures sufficient computational precision is maintained while maximizing memory efficiency, resolving the contradiction between low-precision benefits and precision requirements.
Solution Approach 2:
The patent applies different fixed-point precision configurations to different parts of the CNN architecture based on their specific needs. Critical layers that require higher precision use more bits, while less sensitive layers use fewer bits, optimizing the balance between memory storage and computational precision locally throughout the network.
Data Source
AI summary
A system for operating a floating-to-fixed arithmetic framework includes a floating-to-fix arithmetic framework on an arithmetic operating hardware such as a central processing unit (CPU) for computing a floating pre-trained convolution neural network (CNN) model to a dynamic fixed-point CNN model. The dynamic fixed-point CNN model is capable of implementing a high performance convolution neural network (CNN) on a resource limited embedded system such as mobile phone or video cameras.