Test: Pretrained vs Scratch + Curriculum Learning with Confidence Score

[yeontachi]

Description

##Purpose
This experiment aims to compare the performance of a publicly available pretrained models and analyze its effcet on training performance and generalization

Model & Dataset

• Models : MobileNetV2, ResNet18
• Dataset CIFAR-10
• Evaluation Metrics: Accuracy, Loss, Confidence Score Distribution

Curriculum Learning Strategy

• Initial training uses high-confidence samples
• Gradually introduces lower-confidence (more difficult) samples
• The goal is to simulate a "learning progression" similar to human education

Expected Outcome

• Determine whether a scratch-trained model with curriculum learning can match or outperform the pretrained model
• Evaluate the impact of confidence-based sample scheduling on training stability and generalization
• Analyze the trade-offs between transfer learning and confidence-driven curriculum learning

CheckList

• Evaluate pretrained MobileNetV2, ResNet18 performance on the chosen dataset
• Train MobileNetV2 from scratch and evaluate its performance
• Implement confidence score calculation for each sample
• Design curriculum schedule based on confidence levels
• Train pretrained model with curriculum learning
• Train Scratch model with curriculum learning
• Visualize training/test accuracy and loss curves
• Summarize results in a performance comparison table
• Write detailed analysis and discussion of the results