AI Processor Compressing Operation Data via Channel Segmentation
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing electronic apparatuses face challenges in efficiently compressing operation data from artificial intelligence (AI) models, particularly due to mismatched basic units for CNN algorithms and compression processing, leading to excessive hardware resource usage and transmission bandwidth requirements.
Innovation Solution
An electronic apparatus and method that compress operation data based on the number of non-zero data values from AI model layers, using various coding modes to reduce data size, and store and decompress data within internal memory to optimize hardware resource usage and transmission efficiency.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Quantity of substance
If conventional compression algorithms (JPEG, JPEG2000, PNG, MPEG, Lempel-Ziv) are applied to compress multi-channel feature map data, then storage capacity is reduced, but the basic unit for CNN algorithm processing and compression processing do not coincide, leading to increased device complexity and processing overhead
Solution Approach 1:
The patent segments the feature map data into multiple channels, where each channel is independently compressed using compression algorithms. This segmentation allows the compression basic units to align with the CNN processing units, resolving the mismatch between compression and processing granularities while reducing overall storage requirements.
Solution Approach 2:
The patent changes the parameter of data representation by transforming multi-channel feature map data into a format suitable for compression, applying different compression algorithms to different channels based on their characteristics, and then reconstructing the data for CNN processing. This parameter transformation enables efficient compression without compromising processing efficiency.
2Loss of energy
If operation data is compressed and stored in internal memory, then transmission bandwidth is reduced, but additional compression and decompression processing is required, increasing processing time
Solution Approach 1:
The patent applies compression to operation data before it is stored in internal memory or transmitted between layers. By performing compression in advance (pre-compression), the system reduces the amount of data that needs to be transmitted and stored, thereby reducing transmission bandwidth consumption without significantly impacting overall processing time.
Data Source
AI summary
An electronic apparatus includes a memory storing information on an artificial intelligence (AI) model comprising a plurality of layers, and a processor configured to obtain an output image that is processed from an input image using the AI model. The processor is configured to, based on a number of non-zero data values included in operation data output from a first layer among the plurality of layers, compress the operation data according to at least one of a plurality of coding modes and store the compressed data in an internal memory, obtain restoration data corresponding to the operation data by decompressing the compressed data stored in the internal memory, and provide the obtained restoration data to a second layer among the plurality of layers.