Neural Network Encoding With Mixed Bit Depths for Data Compression
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing neural network-based data compression methods face challenges in efficiently handling diverse data types and meeting footprint constraints while maintaining high compression ratios and image quality.
Innovation Solution
A method involving a neural network model that converts input data into vector data with multiple bit depths using different channel groups, merging and quantizing partial vector data to generate encoded data, and employing multiple encoders and decoders for efficient data processing.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If a single bit depth is used for encoding all vector data, then the encoding process is simple, but compression efficiency is reduced
Solution Approach 1:
The patent divides the vector data into multiple portions (first portion and second portion) and applies different bit depths to each portion. This segmentation allows the system to optimize compression efficiency by using higher bit depth for important data portions and lower bit depth for less critical portions, thereby improving overall compression efficiency without requiring uniformly high bit depth across all data.
Solution Approach 2:
The patent applies different bit depths to different portions of the vector data based on their importance or characteristics. The first portion uses a first bit depth while the second portion uses a second bit depth, allowing each portion to be encoded with the appropriate quality level for its specific requirements, thus optimizing the balance between compression efficiency and data fidelity.
2Productivity
If multiple encoders with different architectures are used, then compression performance is improved, but device footprint increases
Solution Approach 1:
The patent employs a dynamic encoder selection mechanism where the selector neural network chooses between multiple encoders based on the characteristics of the input data. This dynamic approach allows the system to adaptively select the most appropriate encoder for each specific data type or scenario, improving compression performance while avoiding the need to permanently instantiate all possible encoder architectures, thus managing device footprint more effectively.
3Measurement precision
If higher bit depth is used for all vector data, then image quality is improved, but power consumption increases
Solution Approach 1:
The patent applies different bit depths to different portions of the vector data rather than using a uniform high bit depth for all data. By allocating higher bit depth only to critical portions where quality is most important and lower bit depth to less critical portions, the system maintains acceptable image quality while significantly reducing the overall computational complexity and power consumption associated with high-precision encoding.
Solution Approach 2:
The patent uses higher bit depth for only the necessary portions of the data (first portion) rather than applying high bit depth universally. This partial application of high precision encoding maintains quality where it matters most while avoiding the excessive power consumption that would result from applying high bit depth to all data portions.
Data Source
Figure 1
Figure 2
Figure 3
AI summary
A method of processing data using a neural network model, including: receiving, by a first processor, input data; obtaining, by the first processor, first vector data in which the input data is encoded, wherein the first vector data is obtained from a first neural-network-based encoder by providing the input data as an input to the first neural-network-based encoder; converting, by the first processor, a first portion of the first vector data into first partial vector data having a first bit depth; converting, by the first processor, a second portion of the first vector data into second partial vector data having a second bit depth different from the first bit depth; and generating, by the first processor, encoded data based on the first partial vector data and the second partial vector data.