Predictable Area Detection for Neural Network Inference Efficiency

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

In machine learning using convolutional neural networks for image processing, the inclusion of predictable areas in image data leads to increased computation and decreased efficiency and accuracy, as these areas do not contribute significantly to the learning process.

Innovation Solution

An inferring apparatus that determines predictable areas in input data and processes only the non-predictable areas through a machine learning model, reducing computation time and improving accuracy by skipping or simplifying processing of predictable areas.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Measurement precision

If all image data including predictable areas is input to the machine learning model, then complete image processing is achieved, but computation amount increases and processing efficiency decreases

Engineering Contradiction:
Improveinferring accuracyVSAvoidprocessing efficiency
Core Design Contradiction:
Measurement precisionVSProductivity

Solution Approach 1:

The image data is segmented into predictable areas and non-predictable areas. The predictable area determining unit divides the input image into regions that can be easily predicted (such as uniform color regions) and regions that require machine learning processing. This segmentation allows the system to apply different processing strategies to different parts of the image, improving overall efficiency while maintaining accuracy for complex regions.

Inventive Principle:
Principle #1Segmentation

Solution Approach 2:

The predictable area determining unit extracts and identifies predictable areas from the input image data. By taking out these predictable regions, the system prevents them from being processed by the computationally intensive machine learning model, thereby reducing the computation amount and improving processing efficiency while preserving the non-predictable areas for accurate inference.

Inventive Principle:
Principle #2Taking out (Extraction)

2Reliability

If all image data is processed through the machine learning model, then comprehensive inferring is achieved, but computation amount increases

Engineering Contradiction:
Improveinferring accuracyVSAvoidcomputation load
Core Design Contradiction:
ReliabilityVSUse of energy by moving object

Solution Approach 1:

The image is segmented into predictable and non-predictable areas, allowing the machine learning model to process only the necessary portions. This reduces the computation load and energy consumption while maintaining inferring accuracy for the critical non-predictable regions.

Inventive Principle:
Principle #1Segmentation

Solution Approach 2:

Instead of applying the machine learning model to the entire image (excessive action), the system applies it only to non-predictable areas (partial action). This partial processing approach reduces computation load and energy usage while still achieving reliable inferring results for the areas that require it.

Inventive Principle:
Principle #16Partial or excessive action

3Ease of operation

If predictable areas are processed equally with non-predictable areas, then uniform processing is achieved, but learning process efficiency decreases

Engineering Contradiction:
Improveprocessing uniformityVSAvoidlearning efficiency
Core Design Contradiction:
Ease of operationVSProductivity

Solution Approach 1:

The system applies different processing qualities to different areas of the image. Predictable areas receive simple, efficient processing while non-predictable areas receive comprehensive machine learning processing. This local differentiation maintains operational simplicity while dramatically improving learning efficiency by avoiding unnecessary processing of predictable regions.

Inventive Principle:
Principle #3Local quality

Data Source

PatentUS11769221B2Learning apparatus, inferring apparatus, computer program, computer system, inferring method, and learning method
Publication Date: 2023.09.26 AKUSERU KK
  • US11769221B2 patent drawing
  • US11769221B2 patent drawing
  • US11769221B2 patent drawing

AI summary

To provide a learning apparatus and an inferring apparatus that can prevent problems such that a computation amount increases, efficiency of a learning process decreases, and an inferring result does not have high accuracy.A learning apparatus and an inferring apparatus include a predictable area determining unit that determines whether target data has a predictable area in which an inferring result of an inferring process can be easily predicted. In the learning apparatus and the inferring apparatus, predetermined data processing is performed on a data area that is determined to be a predictable area, data required to infer an area that is not a predictable area is output to a machine learning model, and data processing is performed in an average time shorter than that of the inferring process.