Applied Machine Learning

Applied Machine Learning is a hands-on continuation of an introductory AI/ML survey course, moving from paradigms and foundational concepts into the practical work of building, training, and evaluating models. PyTorch is the course’s working tool throughout: every major idea — the perceptron, a neural network layer, a loss function, a gradient step — is introduced as something the learner writes, runs, and inspects, not just reads about.

The course follows the shape of a real applied-ML project. It opens with the perceptron and PyTorch tensors as the mathematical and computational base, then spends two lectures on the less glamorous but decisive work of data retrieval, cleaning, and feature engineering using the Titanic passenger dataset as a running case study. From there it builds up the mechanics of training itself — network architecture, evaluation metrics, loss functions, and gradient-based optimization — before a dedicated lecture on exploratory data analysis as its own disciplined skill. The course closes with two specialized architectures built on everything before them: reinforcement learning (Markov decision processes through Q-learning to deep Q-networks) and transformers (the attention mechanism through pretrained model families and fine-tuning practice).

Throughout, the course treats the learner as someone who already has the degree and now needs the job: questions of AI ethics, bias, and business impact are raised as real tradeoffs a practitioner has to navigate, not settled facts to memorize, and the standard of understanding is “would this hold up in production,” not “would this pass an exam.”

By the end of this course, students will be able to:

1. Implement an end-to-end applied ML pipeline in Pytorch: retrieving, cleaning, encoding, normalizing, and splitting real tabular data before training a model on it.
2. Build the mathematical and architectural foundation of neural networks from a single perceptron up through the shallow vs. deep architecture decision.
3. Select and justify a loss function and an optimization algorighm: SGD, momentum, AdaGrad, RMSprop, or Adam for a given modeling task.
4. Evaluate a trained model with the metric its business context actually calls for and diagnose common training failures.
5. Run a complete exploratory data analysis workflow on an unfamiliar dataset and turn its findings into concrete feature-engineering and modeling decisions.
6. Apply reinforcement learning (Q-learning, deep Q-networks) and transformer-based architectures (attention, pretrained models) to realistic problem framings including the practical work of adopting a pretrained model.

Upon successful completion of this course, students will:

Asse

• Build and Traina PyTorch model on a real tabular dataset end-to-end, from raw file to evaluated result.
• Choose a loss function and optimizerfor a described task and defend the choice against its alternatives.
• Diagnosea described training failure like a vanishing exploding gradient, a stalled loss curve, or an unstable run, and name the practical fix.
• Pick the right evaluation metricfor a business scenario (precision/recall/F1/ROC-AUC for classification; MAE/MSE/R Squared for regression) instead of defaulting to accuracy.
• Analyze a PyTorch snippetlike a tensor operation, a DataLoader, or a training-loop step and predict its output or spot its bug.
• Explainthe Q-learning update rule and the attention formulawithout needing to re-derive it from scratch.
• Discuss AGI, algorithmic fairness, and deepfake detection even-handedly as considerations a practitioner has to weigh.

1. Basics of ML and Data
2. Practical Applications of ML and AI
3. PyTorch and Visualization
4. Concepts of Perceptron
5. Understanding Data, Features, and Notations for Applied ML
6. Data Retrieval, Cleaning, and Wrangling
7. Shallow vs Deep Neural Networks
8. Measuring Performance
9. Loss Functions
10. Gradient Descent
11. EDA: A Practical Guide
12. Reinforcement Learning
13. Transformers