ML_2021_Spring

ML course in NTU: https://speech.ee.ntu.edu.tw/~hylee/ml/2021-spring.html
All inital codes are provided by TAs

• HW1: Regression of COVID by Neural Network
  • Kaggle Leaderboard (Public, Overall / Total): 2, 449 / 2032
  • Overfitting
• HW2: Phoneme Classification from MFCC
  • Kaggle Leaderboard (Public, Overall / Total): 110, 112 / 1522
  • Separable Conv1d
  • Feature Scaling
  • Voting
• HW3: Food Classification with Semi-supervised learning
  • Kaggle Leaderboard (Public, Overall / Total): 10, 46 / 1404
  • AutoAugment
  • Unbiased Teacher like pipeline
  • Soft pseudo-label
  • Voting with data augmentation on testing set
• HW4: Speaker classification by Self-Attention
  • Kaggle Leaderboard (Public, Overall / Total): 448, 545 / 1170
• HW8: Anomaly Detection
  • Kaggle Leaderboard (Public, Overall / Total): 219, 220 / 1193
• HW10: Adversarial Attack
  • MI-FGSM on ensembled model
• HW11: Domain Adaptation from img IRL to sketch
  • Kaggle Leaderboard (Public, Overall / Total): 15, 16 / 1061
  • DANN
  • Data Augmentation
    • Impulse Noise
    • Random erase with either black or white
    • Canny Edge Detector with random threshold
  • Balance the predicted distribution by class-depended weight on loss
  • Class distribution loss by temperature-softmax
  • Multi-task learning with AutoEncoder (reconstruction task)
  • Pseudo label with relative probability
  • Reassign labels to balance class-distribution while inference
• HW13: Network Compression (extend of HW3 but less parameters)
  • Kaggle Leaderboard (Public, Overall / Total): 7, 3 / 590
  • Separable Conv2d instead of Vanilla Conv2d
  • AutoAugment + RandomErasing
  • Dropout