🧠 [2025-05-29] Fine-Tuning the Real-Time Recognizer + CIFAR-10 CNN Practice

Today I focused on two key areas:

1. Continuing CNN architecture studies with CIFAR-10 as a baseline dataset
2. Running two major experiments on the Real-Time Daily Activity Recognizer, which I had fully reset yesterday (2025/05/28)

🧪 1. CIFAR-10 CNN Classifier (Book-based Practice)

As part of my deep learning fundamentals, I replicated a CNN-based CIFAR-10 classifier following a book tutorial.

✅ What I Did

• Loaded CIFAR-10 via keras.datasets
• Preprocessed integer labels with to_categorical
• Constructed a simple CNN with:
  • Conv2D, MaxPooling2D, Flatten, Dense, and Dropout
• Identified and corrected common mistakes like:
  • input_shape=(...) syntax
  • Typos in label variables (y_trian vs y_train)

🧠 Key Takeaways

• Practicing basic architecture helped reinforce layer-wise understanding
• Debugging minor issues greatly improved my attention to syntax and logic
• This exercise provided a useful baseline for evaluating deeper models

🔁 2. Real-Time Activity Recognizer — Model Experiments

After resetting the project yesterday, I ran two structured experiments today aimed at improving validation performance and resolving class confusion (notably with brushing_teeth).

🧪 Experiment A – Fine-Tuning with Class Weighting

• Backbone: MobileNetV2 (last 20 layers unfrozen)
• Augmentation: Moderate (rotation ±30°, channel/brightness shift, etc.)
• Regularization: Dropout(0.5), class_weight applied for walking
• Training: EarlyStopping + ReduceLROnPlateau
• Result:
  • Best Val Acc: 0.5370 at Epoch 9
  • Fine-tuning gave a +10% improvement over frozen backbone
  • Train/Val gap narrowed, training behavior more stable
  • 📈
    

🧪 Experiment B – Addressing Class Confusion

• Focus: Reduce misclassifications to brushing_teeth
• Augmentation: More aggressive (rotation ±45°, zoom 0.4, brightness 0.6–1.4, etc.)
• Same Backbone: MobileNetV2, last 20 layers unfrozen
• Result:
  • Validation accuracy similar (~0.54) but with healthier learning dynamics
  • Overfitting signs reduced, validation loss decreased steadily
  • Misclassification pattern began to improve
  • 📈
    

🧠 Overall Insights

“Refining after a reset works — and clearer logs give better direction.”

• Transfer learning with fine-tuning is significantly better than frozen models
• Class weighting and aggressive augmentation helped reduce class imbalance issues
• Focusing on class-specific misclassification (brushing_teeth) was a valuable strategy

🎯 Next Steps

• 📸 Complete image crawling (aiming for ~3000 images/class)
• 🤖 Try stronger backbones (EfficientNetB0, ResNet50, etc.)
• 🧪 Apply class-specific augmentation
• 🔍 Analyze updated confusion matrices

Even though I started fresh yesterday, today marked the first meaningful progress in the rebuilt pipeline.
Performance is improving steadily — and I’m just getting started.