Independent Projects and Competition Submissions on Data Science, Machine and Deep Learning
-
Titanic Survival Prediction - Kaggle Link
- Position: Top 33% [78% Accuracy]
- Model: Random Forest (/w Parameter gridsearch)
- Data Preprocessing: Missing values and outliers removal (cleaning), replaced age data to categorical bins
-
MNIST Digit Recognizer - Kaggle Link
- Position: Top 18% [94.5% Accuracy] (91% without data augmentation)
- Model: CNN (Modified LeNet, 2Conv, 2FC)
- Data Preprocessing: Normalization and Data Augmentation (Translate and Rotate-Crop)
- Optimizer: SGD + momentum with Step LR decay
- Regularization: Dropout (0.2)
-
CIFAR-10 Image Classification - Kaggle Link
- Position: 81% Accuracy (74% without data augmentation)
- Model: CNN (4 Conv, 3FC)
- Optimizer: Adam with LR decay
- Data Preprocessing: Normalization and Data Augmentation (Translate, Rotate and Zoom)
- Regularization: Dropout (0.25) or BatchNormalisation [Both gave similar performance]
-
Facial Keypoint Detection - Kaggle Link
- Position: 2.456 (Mean Average Error)
- Model: CNN (3 Conv, 3FC)
- Optimizer: Adam with LR decay
- Data Preprocessing: Normalization
- Regularization: Dropout (0.25)