Neural Network Feature Compression for Reliable Inference
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional methods for machine learning inference processing face reliability issues due to varying data sizes and compression methods that can prevent consistent inference, particularly in cases where data size after compression is undefined, leading to unpredictable inference success.
Innovation Solution
An information processing apparatus and method that includes a preliminary analysis unit to generate control information for controlling compression and decoding of feature amounts, using a compression unit to compress and a decoding unit to decode feature amounts based on this control information, ensuring reliable inference processing.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Quantity of substance
If dynamic compression rate is used to reduce data size, then memory usage and data communication amount are reduced, but inference reliability deteriorates because data size after compression cannot be defined
Solution Approach 1:
The system performs preliminary analysis of the neural network structure and data characteristics before compression to determine appropriate compression parameters. This advance preparation ensures that compression settings are optimized for the specific inference task, guaranteeing that compressed data will maintain sufficient quality for reliable inference while achieving the desired size reduction.
Solution Approach 2:
The system dynamically adjusts compression parameters based on analysis results rather than using fixed or purely data-driven compression rates. By changing compression parameters (such as quantization precision, block size, or encoding settings) based on neural network architecture and data statistics, the system achieves both size reduction and inference reliability.
2Volume of stationary object
If compression is applied to reduce data size, then memory capacity requirements are reduced, but data processing reliability deteriorates due to undefined compressed data size
Solution Approach 1:
The system performs preliminary analysis of the neural network structure and data characteristics before compression to determine appropriate compression parameters. This advance preparation ensures that compression settings are optimized for the specific inference task, guaranteeing that compressed data will maintain sufficient quality for reliable inference while achieving the desired size reduction.
Solution Approach 2:
The system incorporates feedback mechanisms where compression parameters are adjusted based on the actual performance of inference processing. By monitoring whether compressed data maintains sufficient quality for accurate inference and using this feedback to refine future compression settings, the system ensures both memory efficiency and processing reliability.
3Quantity of substance
If dynamic compression is used, then data size is reduced, but the complexity of controlling compression and decoding increases
Solution Approach 1:
The system performs preliminary analysis of the neural network structure and data characteristics before compression to determine appropriate compression parameters. This advance preparation ensures that compression settings are optimized for the specific inference task, guaranteeing that compressed data will maintain sufficient quality for reliable inference while achieving the desired size reduction.
Data Source
Figure 1~2
Figure 3
Figure 4A~4B
AI summary
The present disclosure relates to an information processing apparatus and an information processing method that enable inference of machine learning to be performed more reliably. Provided is an information processing apparatus including: a preliminary analysis unit that analyzes a neural network before inference using the neural network is performed and generates control information for controlling compression and decoding of a feature amount processed by the neural network; a computing unit that performs inference using input data and the neural network and outputs a processing result obtained by processing the feature amount obtained from the input data as a computing result; a compression unit that compresses the feature amount input from the computing unit on the basis of the control information and records a compressed feature amount that is the feature amount having been compressed in a memory; and a decoding unit that decodes the compressed feature amount temporarily recorded in the memory on the basis of the control information and outputs the decoded feature amount to the computing unit. The present disclosure can be applied to, for example, a machine learning inference unit.