Neural Vector Quantization for Footprint-Limited SoC Compression
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing neural network-based data compression and reconstruction technologies face challenges in efficiently managing bit depths and compression ratios while adhering to footprint constraints, particularly in systems-on-chip environments.
Innovation Solution
A method involving a first processor that converts input data into vector data with varying bit depths using multiple neural-network-based encoders and decoders, generating encoded data through bit clamping and merging, and employing a neural-network-based selector to select appropriate decoders for reconstruction.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Ease of manufacture
If uniform bit depth is used for all vector data, then implementation is simple, but compression efficiency and image quality are limited
Solution Approach 1:
The patent applies local quality by using different bit depths for different portions of vector data based on their importance. Specifically, it divides vector data into multiple portions and applies first bit depth to a first portion and second bit depth to a second portion, where the bit depths are different. This allows important regions to maintain higher precision while less important regions use lower precision, thereby improving compression efficiency without significantly degrading overall image quality.
2Measurement precision
If higher bit depth is used, then image quality is improved, but data size and bandwidth requirements increase
Solution Approach 1:
The patent divides vector data into multiple portions and applies different bit depths to different portions. Important regions (first portion) are encoded with first bit depth to maintain high image quality, while less important regions (second portion) are encoded with second bit depth to reduce data size. This selective approach optimizes the balance between image quality and data size by allocating bit depth resources according to regional importance.
3Productivity
If multiple neural network encoders with different bit depths are used, then compression performance is improved, but device complexity increases
Solution Approach 1:
The patent segments the encoding process by using multiple neural network encoders, where each encoder is configured with a specific bit depth. The first neural network encoder uses first bit depth and the second neural network encoder uses second bit depth. This segmentation allows the system to process different portions of vector data with appropriate precision levels, improving overall compression performance while managing complexity through modular encoder design.
4Quantity of substance
If bit depth is reduced for compression, then bandwidth usage is optimized, but reconstruction accuracy deteriorates
Solution Approach 1:
The patent applies different bit depths to different portions of vector data to balance bandwidth usage and reconstruction accuracy. Important regions are encoded with higher bit depth to maintain reconstruction accuracy, while less important regions use lower bit depth to reduce bandwidth requirements. This selective bit depth assignment ensures that bandwidth is optimized without significantly compromising the accuracy of critical regions.
Data Source
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.