Sparse Data Stream Decoding Without Full Matrix Reconstruction
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing data compression methods, particularly for large volumes of data such as audio and video streams and Deep Neural Networks, face high computational complexity and resource constraints, especially in environments with limited hardware and software resources, leading to prohibitive inference complexity.
Innovation Solution
A method for compressing and decompressing data using a compressed sparse format, where the compressed sparse format is directly computed during entropy decoding, bypassing the need to reconstruct the entire matrix, thereby reducing memory requirements and inference complexity by utilizing CABAC technology to extract zero/non-zero coefficients directly from the bitstream.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Quantity of substance
If traditional compression methods are used to compress large volume data such as video streams and DNN parameters, then compression can be achieved, but computational complexity and resource requirements become prohibitively high
Solution Approach 1:
The patent extracts and processes only the significant non-zero coefficients from the transformed video data, rather than encoding all coefficients. By identifying and separately handling the sparse non-zero elements, the system reduces the amount of data that requires complex compression processing while maintaining reconstruction quality.
Solution Approach 2:
The patent segments the compression process into distinct stages: transformation to identify non-zero coefficients, separate encoding of significant coefficients, and deferred reconstruction. This segmentation allows computationally intensive operations to be performed only on necessary data portions, reducing overall computational complexity.
2Reliability
If full data reconstruction is performed during decompression, then complete data recovery is achieved, but memory requirements and processing time increase significantly
Solution Approach 1:
The patent performs preliminary encoding of significant coefficients during the compression phase, organizing data in a format that enables direct reconstruction of only necessary elements. This preliminary organization allows the decoder to skip full reconstruction steps and directly access or compute only the required data portions during inference.
Solution Approach 2:
The patent implements partial reconstruction by computing or retrieving only the significant non-zero coefficients needed for the specific inference task, rather than fully reconstructing the entire dataset. This partial action approach maintains sufficient accuracy for inference while dramatically reducing memory and time requirements.
3Productivity
If existing compression schemes are applied to DNN parameters, then data transmission efficiency improves, but inference complexity remains prohibitive on resource-constrained devices
Solution Approach 1:
The patent extracts and transmits only the significant non-zero DNN parameters rather than all parameters. By identifying and separately encoding the sparse non-zero elements of the transformed DNN weights, the system reduces transmission data volume while enabling more efficient inference on resource-constrained devices through direct use of the compressed representation.
Data Source
AI summary
Procedures, methods, architectures, apparatuses, systems, devices, and computer program products, according to a first aspect, for compressing data including encoding at least one information representative of a use, during the compression, of a compressed sparse format.Procedures, methods, architectures, apparatuses, systems, devices, and computer program products, according to a second aspect, for decompressing input data comprising obtaining information representative of zero or non-zero values in at least a part of the input data, and using only the non-zero values of the zero or non-zero values for a further processing of the part of the input data, based on the representative information.