Ensemble Neural Network Recognition via Feature Vector Segmentation
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current recognition methods using neural networks face challenges in effectively combining the predictions of multiple models with the same structure to enhance recognition accuracy and efficiency, particularly in processing input data such as images, audio signals, and text.
Innovation Solution
The method involves determining a feature vector from input data using an ensemble model comprising two neural networks with the same structure, where connection weights between layers are used to generate partial feature vectors, and these are combined to recognize input data, allowing for parallel implementation and efficient resource utilization.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If multiple neural network models are used to enhance recognition accuracy, then recognition accuracy is improved, but device complexity increases
Solution Approach 1:
The patent segments the neural network computation into distinct phases: an encoding phase that processes input data through encoder layers to generate encoded representations, and a decoding phase that processes multiple model predictions through decoder layers to produce final recognition results. This segmentation allows multiple models to be processed in an organized manner, managing complexity while maintaining accuracy improvements.
Solution Approach 2:
The patent merges the outputs of multiple neural network models by combining their predictions in the decoding phase. The decoder layers integrate results from multiple encoder-decoder models, effectively combining their recognition capabilities to achieve improved accuracy while managing the complexity of having multiple models through a unified processing framework.
2Measurement precision
If multiple models are processed sequentially, then recognition accuracy is improved, but processing time increases
Solution Approach 1:
The patent implements periodic action by alternating between encoding operations (processing input data through encoders) and decoding operations (combining model predictions through decoders). This periodic structure organizes the processing of multiple models into rhythmic phases, allowing systematic integration of multiple model outputs while managing processing time through structured alternation between computation stages.
3Adaptability or versatility
If models with different structures are used, then adaptability is improved, but device complexity increases
Solution Approach 1:
The patent implements universality by using identical encoder and decoder structures that can process inputs from multiple different neural network models. The encoder-decoder framework serves as a universal processing architecture that handles various model outputs through the same computational layers, allowing adaptability to different model types while maintaining structural simplicity and avoiding the need for model-specific processing paths.
Data Source
AI summary
A method of recognizing input data includes determining a feature vector corresponding to an ensemble model from input data, based on the ensemble model, and recognizing the input data based on the feature vector. The ensemble model includes a first model and a second model having a structure that is the same as a structure of the first model.