Partial Discrimination Calculation for Neural Network Memory Reduction
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing image processing apparatuses using hierarchical neural networks require large-capacity memory for feature detection, leading to increased memory costs and inefficiencies.
Innovation Solution
A discrimination calculation apparatus and method utilizing a Convolutional Neural Network (CNN) with a partial discrimination calculation process unit that performs linear discrimination using intermediate hierarchical layer data, reducing the storage capacity needed for feature plane data and implementing data reduction techniques to minimize memory usage.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If a large-scale hierarchical neural network is used for feature detection, then detection precision is improved, but memory capacity requirements increase
Solution Approach 1:
The patent segments the hierarchical neural network computation into multiple stages, storing intermediate results at each layer. This allows the system to process large-scale networks by breaking down the computation into manageable segments that can be stored and processed sequentially, reducing the peak memory capacity required while maintaining detection precision.
Solution Approach 2:
The patent performs preliminary computation and stores intermediate hierarchical layer results before final discrimination. By pre-calculating and storing feature plane data from intermediate layers, the system avoids re-computation and reduces the memory capacity needed for the final discrimination stage, while preserving the benefits of the large-scale network.
2Measurement precision
If all intermediate hierarchical layer data is stored for discrimination calculation, then discrimination accuracy is improved, but storage capacity requirements increase
Solution Approach 1:
The patent extracts and stores only the essential intermediate hierarchical layer results that are most critical for discrimination accuracy. By selectively extracting key feature plane data from intermediate layers rather than storing all possible intermediate results, the system maintains discrimination accuracy while significantly reducing the storage capacity required.
Solution Approach 2:
The patent applies different storage strategies to different hierarchical layers based on their importance for discrimination. Critical intermediate layers are stored with higher fidelity, while less critical layers use compressed or selective storage, optimizing the balance between discrimination accuracy and storage capacity requirements.
Data Source
Figure 1
Figure 2
Figure 3
AI summary
In order to perform a discrimination calculation using a small-capacity storage unit, feature calculation means configured to sequentially calculate a feature of discrimination target data for each hierarchical layer, discrimination calculation means configured to sequentially perform a partial discrimination calculation on the discrimination target data using the feature sequentially calculated by the feature calculation means and store a result of the partial discrimination calculation in a discrimination result storage unit, and control means configured to control the discrimination calculation means to perform a next partial discrimination calculation using the feature sequentially calculated by the feature calculation means and the result of the partial discrimination calculation stored in the discrimination result storage unit and to store a result of the next partial discrimination calculation in the discrimination result storage unit.