Application Research for Fusion Model of Pseudolabel and Cross Network

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Deep convolutional neural networks often struggle with missing supervised information and limited generalization.
• Existing methods face challenges in effectively utilizing unlabeled data and improving model robustness.

Purpose of the Study:

• To address the limitations of supervised information scarcity and weak generalization in deep convolutional neural networks.
• To propose a novel fusion model combining pseudolabeling (PL) from Weakly Supervised Learning (WSL) and Cross Network (CN) from Multitask Learning (MTL).

Main Methods:

• Utilized pseudolabeling to predict and generate labels for data with missing supervised information.
• Employed Cross Network multitask learning, training with both pseudolabeled and labeled data as separate tasks.
• Iteratively refined the model by selecting optimal models based on accuracy and generalization, then using them to generate new pseudolabels.

Main Results:

• The fusion model demonstrated suitability for multitask training across diverse datasets.
• Achieved a 64.76% accuracy for the main task of Facial Beauty Prediction, surpassing conventional methods.
• Showcased improved generalization ability and effective handling of datasets with missing supervised information.

Conclusions:

• The proposed pseudolabeling and Cross Network multitask learning fusion model effectively enhances deep learning performance.
• This approach offers a viable solution for datasets with limited labeled data and weak generalization capabilities.
• The model shows promise for various applications, including Facial Beauty Prediction, by leveraging multitask learning.