AI Encoding and Decoding Apparatus Using Memory-Efficient Integer Parameters
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Deep Neural Networks (DNNs) require significant memory and computational resources for convolution operations, especially when handling high-resolution images, due to the need to store and process real-number parameters and intermediate results, which leads to inefficiencies in memory usage and calculation time.
Innovation Solution
The implementation of memory-efficient integer values for parameters and intermediate results in DNNs, achieved through joint training of AI encoding and decoding networks, where parameters are determined as integer values associated with each other, reducing memory and calculation requirements while minimizing conversion errors.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If real-number parameters are used in convolution layers of DNN, then calculation precision is maintained, but memory usage and calculation time increase significantly
Solution Approach 1:
The patent changes the data type parameter from real numbers to fixed-point numbers (Q-format) for representing convolution parameters and intermediate results. This parameter change reduces memory usage while maintaining sufficient calculation precision for image processing tasks.
Solution Approach 2:
The patent uses simplified fixed-point number representations instead of complex real-number floating-point formats, creating a lighter, more efficient data structure that consumes less memory and processing resources while achieving comparable results.
2Measurement precision
If real-number parameters are used in convolution layers of DNN, then calculation precision is maintained, but calculation time increases
Solution Approach 1:
The patent changes the computational parameter representation from floating-point real numbers to fixed-point numbers, which enables faster arithmetic operations while maintaining adequate precision for convolution operations in image processing.
3Reliability
If intermediate result values are stored with high precision, then conversion errors are minimized, but memory requirements increase
Solution Approach 1:
The patent applies fixed-point quantization parameters to intermediate result values, changing their representation from high-precision real numbers to compact fixed-point formats. This reduces memory requirements while controlling conversion errors through appropriate quantization strategies.
Data Source
AI summary
An artificial intelligence (AI) decoding apparatus obtains image data corresponding to a first image, which is AI-downscaled from an original image by an AI encoding apparatus by using a first deep neural network (DNN); reconstructs a second image corresponding to the first image, based on the image data; and obtain a third image, which is AI-upscaled from the second image, convolution is performed based on the second image and second parameters of a filter kernel included in a second DNN, wherein each of the second parameters is an integer value, and the second parameters are determined as values associated with first parameters of a filter kernel included in the first DNN. Embodiments use memory-efficient values with respect to filter kernels. Parameters used to obtain the memory-efficient integer values may be obtained via joint training between the first DNN and the second DNN.