Machine Learning & Deep Learning Implementations

A comprehensive archive of foundational to advanced Machine Learning, Deep Learning, and Reinforcement Learning algorithms. This repository emphasizes building models from the ground up, understanding mathematical mechanics, and systematic performance tuning and evaluation across various datasets.

---

📚 Master Table of Projects & Implementations

1. Supervised Learning: Regression & Classification

Project / Domain	Implemented Algorithms	Key Tech / Methods	Summary of Results	View
Multi Dataset Regression	Logistic Regression, Linear, Logistic, ElasticNet	EDA, L1/L2 , Gradient Descent, OLS, IQR Outlier Removal, One-Hot Encoding	Achieved 96.0% accuracy on Penguins classification; successfully prevented overfitting via L1/L2 penalties	🔗 Code

2. Deep Learning: Computer Vision & Vision Models

Project / Domain	Implemented Algorithms	Key Tech / Methods	Summary of Results	View
PyTorch Neural Networks to VGG-13	Custom NN, CNN, VGG-13	PyTorch from scratch, Batch Norm, Early Stopping, LR Scheduling, EMNIST (36 classes)	Achieved 92.0% accuracy on EMNIST test set; demonstrated trade-offs of adapting VGG-13 to $28 \times 28$ image constraints	🔗 Code
ResNet-18 vs. VGG-16 Ablation & Vanishing Gradients	ResNet-18, VGG-16 (Ver. C), VGG-Deep	Residual Skip Connections, He vs. Xavier Init, SGD vs. Adam, L2 Gradient Norm Logging, 64x64 RGB Dataset (3 classes)	ResNet-18 achieved 98.75% accuracy (vs. VGG-16 at 92.37%); experimentally proved severe vanishing gradients in deep VGG layers via L2 norm tracking	🔗 Code
Air Quality Time-Series Forecasting	Stacked 3-Layer LSTM, Unidirectional LSTM	AirQualityUCI Dataset, Missing Value Interpolation (-200 Spikes), BPTT, Epoch & LR Sensitivity Optimization	Achieved R² of 0.80 and MAE of 0.248; demonstrated unidirectional LSTMs outperform bidirectional setups for forward hourly regression	🔗 Code

3. Reinforcement Learning

Project / Domain	Implemented Algorithms	Key Tech / Methods	Summary of Results	View
MDP to n-step Double Q-learning	Custom MDP (Gymnasium), SARSA (On-Policy), n-step Double Q-Learning	Temporal-Difference Control, Safety in AI Protocols, $\epsilon$-decay Tuning, Two-Q-table Bias Mitigation	SARSA achieved superior stability and faster convergence compared to n-step Double Q-learning ($n=3$); implemented boundary safety validation without external RL solvers	🔗 Code

---

📁 Repository Structure

ml-algorithm-implementations/
├──01-supervised-learning/
│   └── multi-dataset-regression/
│       ├── 01_data_preprocessing_and_eda.ipynb
│       ├── 02_logistic_regression.ipynb
│       ├── 03_ols_linear_and_ridge_regression_wine.ipynb
│       └── 04_elasticnet_gradient_descent_diamonds.ipynb
├──02-deep-learning/
│   ├── pytorch-neural-networks-to-vgg13
│   │   ├── 01_binary_classification_nn.ipynb
│   │   ├── 02_binary_classification_nn_optimization.ipynb
│   │   ├── 03_emnist_cnn_classification.ipynb
│   │   └── 04_emnist_vgg13_adaptation.ipynb
│   ├── cnn-resnet18-vs-vgg16-ablation/
│   │   ├── 01_resnet18_vs_vgg16_ablation_study.ipynb
│   │   └── 02_vanishing_gradient_l2_norm_verification.ipynb
│   └── air-quality-lstm-forecasting
│       └── air_quality_lstm_time_series_forecasting.ipynb
├──03-reinforcement-learning
│   └── mdp-to-nstep-double-q-learning
│       ├── 01_environment_definition_and_safety_test.ipynb
│       ├── 02_sarsa_tabular_implementation.ipynb
│       └── 03_nstep_double_q_learning_analysis.ipynb