Split Layer Classification Model for Labeling Error Mitigation
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing classification models in artificial intelligence systems often produce erroneous results due to incorrect labeling of learning data, leading to reduced classification accuracy, as they rely on human-labeled data that may contain mistakes.
Innovation Solution
An electronic device is configured to split a feature vector obtained via a convolution layer into designated subfeatures, using split layers for learning, which allows for more accurate classification by adjusting parameters based on labeling information and reducing overfitting.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If learning data is labeled by human beings, then the classification model can be trained, but mistakes may cause wrong labels to be assigned leading to reduced classification accuracy
Solution Approach 1:
The patent segments the learning data into multiple groups and assigns different weights to each group. By dividing the data and applying differential weighting, the system can reduce the impact of labeling errors in certain segments while maintaining the overall training effectiveness, thereby improving classification accuracy despite the presence of harmful labeling mistakes.
Solution Approach 2:
The patent introduces weight parameters that can be adjusted to compensate for labeling errors. By changing the weight parameters assigned to different learning data based on their reliability or source, the system optimizes the training process to minimize the impact of wrong labels and improve the final classification accuracy.
2Reliability
If a classification model processes all input data uniformly, then the model structure remains simple, but it cannot effectively handle noisy environments or complex feature combinations
Solution Approach 1:
The patent segments the learning data into multiple groups and applies different processing weights to each segment. This segmentation allows the model to handle different types of data (including noisy data) differently, improving robustness without requiring a fundamentally more complex model architecture.
Solution Approach 2:
The patent introduces adjustable weight parameters that control how different learning data groups contribute to the training process. By dynamically adjusting these parameters, the model can adapt to noisy environments and complex feature combinations while maintaining a relatively simple overall structure.
Data Source
AI summary
According to an embodiment, an electronic device comprises at least one processor and a memory, wherein the memory stores instructions that, when executed, cause the at least one processor to control the electronic device to: obtain data to be classified, obtain a feature vector from the data by performing convolution on the data and a plurality of filters using a classification model stored in the memory, identify outputs corresponding to subfeatures using a split layer including the subfeatures resulting from splitting the feature vector, and output a class corresponding to the data based on the outputs.