University AI Course

This repository contains the complete set of practical assignments developed for the Artificial Intelligence course at the university. The work covers a wide spectrum of machine learning and deep learning paradigms — from classical supervised learning and ensemble methods to generative models and recurrent neural networks for time-series forecasting.

Each assignment is structured as a fully functional Jupyter notebook, complete with data preprocessing, model implementation, hyperparameter optimisation, performance evaluation, and visualisation. Together, they demonstrate a solid command of the theoretical foundations and practical implementation skills required in modern AI.

---

Course Context & Learning Objectives

These assignments were designed to bridge the gap between theoretical AI concepts and hands-on implementation. The key learning outcomes addressed include:

• Implementing and evaluating classic ML algorithms (KNN, SVM, Gradient Boosting, XGBoost).
• Understanding and applying transfer learning and generative adversarial networks (GANs).
• Mastering deep learning for time-series using RNNs, LSTMs, and GRUs.
• Performing hyperparameter tuning, model interpretation, and feature importance analysis.
• Working with real‑world datasets and image data.

---

Repository Structure

Notebook	Topic	Key Techniques
assignment1.ipynb	Supervised Learning & Ensemble Methods	KNN, SVM, Gradient Boosting, XGBoost, GridSearchCV, SHAP, Feature Importance
assignment2.ipynb	Deep Learning for Vision – GANs & Transfer Learning	DCGAN, VGG16, Transfer Learning, Data Augmentation, Fine‑tuning
assignment3.ipynb	Time Series Forecasting with RNNs	RNN, LSTM, GRU, Bidirectional RNNs, Early Stopping, Learning Rate Scheduling

---

Assignment 1 – Supervised Learning & Model Interpretation

Objective

Compare multiple classifiers on three benchmark datasets, optimise hyperparameters using cross‑validation, and interpret the best model using SHAP values.

Datasets

• Iris – 3‑class flower classification (150 samples).
• Titanic – binary survival prediction (891 samples, pre‑processed).
• Telco Churn – binary customer churn prediction (7,043 samples).

Models & Implementation

• K‑Nearest Neighbours (KNN): Tuned number of neighbours and distance metrics.
• Support Vector Machine (SVM): Used RBF kernel; tuned C and gamma parameters via GridSearchCV.
• Gradient Boosting: Built with GradientBoostingClassifier, optimised learning rate and tree depth.
• XGBoost: The highlight of this assignment. Implemented with:
  • GridSearchCV over max_depth, learning_rate, reg_lambda, and scale_pos_weight (to handle class imbalance).
  • Early stopping and evaluation on validation set.
  • SHAP (SHapley Additive exPlanations): Generated beeswarm plots to identify the most influential features for churn prediction.

Results

• Achieved >80% accuracy on Titanic and Telco churn; >95% on Iris.
• SHAP analysis provided transparent, business‑ready insights into customer churn drivers.

---

Assignment 2 – Deep Learning for Vision Tasks

Part A – Generative Adversarial Network (GAN) on CIFAR‑10

• Designed and trained a DCGAN (Deep Convolutional GAN) from scratch using Keras.
• Generator: Built with Conv2DTranspose layers and LeakyReLU activations to upsample a 100‑dim latent vector into a 32×32 RGB image.
• Discriminator: A CNN with dropout (0.4) and LeakyReLU to distinguish real from synthetic images.
• Trained for 5 epochs with batch size 1024, showing stable convergence and realistic sample generation.

Part B – Transfer Learning for Binary Image Classification (Cats vs Dogs)

• Leveraged a pre‑trained VGG16 (ImageNet weights) as a feature extractor.
• Data augmentation (rotation, shift, shear, zoom, flip) was applied to the training set to mitigate overfitting.
• Fine‑tuned the top layers with dense layers and dropout.
• Achieved >90% validation accuracy in only 5 epochs, demonstrating the efficiency of transfer learning on small datasets.

Key Takeaways

• Hands‑on experience with adversarial training dynamics.
• Effective use of pre‑trained CNNs and data augmentation for limited‑data scenarios.

---

Assignment 3 – Time Series Forecasting with RNNs

Objective

Predict 72 future temperature readings (10‑minute intervals) given 216 past readings from the Jena Climate dataset.

Models Implemented

• Simple RNN: Baseline model with 1–3 layers.
• LSTM: Deep LSTM with dropout to handle long‑term dependencies.
• GRU: Faster alternative to LSTM with comparable performance.
• Bidirectional GRU (final model): Two bidirectional GRU layers with L2 regularisation and dropout.

Preprocessing Pipeline

• Scaled data using MinMaxScaler.
• Created sliding‑window sequences: 216 input → 72 output.
• Split data 90‑10 into training and validation sets.

Optimisation

• Used EarlyStopping to prevent overfitting.
• Applied ReduceLROnPlateau for dynamic learning rate adjustment.
• Final model achieved R² = 0.94 on the test set (a significant improvement over the baseline RNN which scored R² = 0.70).

Output

• Plots comparing predictions against actual values.
• Loss curves for training and validation sets.

---

Technical Stack

• Language: Python 3.10+
• Libraries: TensorFlow, Keras, Scikit‑learn, XGBoost, Pandas, NumPy, Matplotlib, Seaborn, SHAP, OpenCV
• Environment: Jupyter Notebook, Google Colab

---

How to Run

1. Clone this repository:

   git clone https://github.com/yourusername/ai-course-assignments.git
   cd ai-course-assignments

2. Install required packages:

   pip install -r requirements.txt

3. Launch Jupyter and open the desired assignment:

   jupyter notebook assignment1.ipynb

Note: Some datasets (e.g., skin lesion images, Jena climate CSV) are not included. Please download them separately and update file paths. The CIFAR‑10 dataset is built into Keras and loads automatically.