AI Encoding and Decoding Apparatus with DNN Normalization
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Deep neural networks (DNNs) require significant memory and computational resources for convolution operations, especially in high-resolution image processing, leading to increased transformation errors when using low-precision transformations to reduce memory needs.
Innovation Solution
An AI encoding and decoding apparatus that employs a second DNN trained in conjunction with a first DNN to minimize transformation errors by normalizing and clipping result values into integer ranges, using scale factors to maintain distribution equality and reduce memory requirements.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Quantity of substance
If low-precision transformation is used to reduce memory requirements, then memory usage is reduced, but transformation error increases
Solution Approach 1:
The patent applies parameter changes by transforming intermediate result values from floating-point format to fixed-point format with specific precision (e.g., 8-bit or 16-bit integers). This parameter transformation reduces memory usage while maintaining acceptable transformation accuracy through careful selection of precision levels and range scaling.
Solution Approach 2:
The patent introduces scale factors as intermediary elements that mediate between high-precision computation requirements and low-precision storage constraints. These scale factors enable accurate representation of transformed values in reduced precision by adjusting the range and distribution of quantized values.
2Productivity
If convolution operations are performed in multiple layers, then processing capability is improved, but memory and calculation requirements increase
Solution Approach 1:
The patent applies parameter changes by converting intermediate result values from high-precision floating-point format to low-precision fixed-point format between convolution layers. This parameter transformation significantly reduces memory requirements for storing intermediate results while maintaining sufficient processing capability through careful precision management.
Solution Approach 2:
The patent segments the computation process into distinct stages with precision conversion points between convolution layers. By dividing the deep neural network processing into segments with intermediate quantization, the patent reduces cumulative memory requirements while preserving essential processing capabilities across multiple layers.
Data Source
AI summary
Provided is an artificial intelligence (AI) decoding apparatus including a memory storing one or more instructions; and a processor configured to execute the one or more instructions to, when an image is input to a second DNN including a plurality of layers, obtain first result values based on an operation between the image and a first filter kernel and obtain second result values based on an operation between the image and a second filter kernel, from a first layer including the first and second filter kernels from among the plurality of layers, perform normalization by transforming the first result values into first values by using a first scale factor, and, perform normalization by transforming the second result values into second values by using a second scale factor, transform the first values and the second values into integer values included in a preset range.