AI academy- From Decision Boundaries to Predictive Modelling with Neural Networks

This workshop provides a practical bridge from classical supervised machine learning to predictive modelling with feed-forward neural networks.

The morning begins with the idea of a decision boundary: how a model partitions feature space in order to make a prediction. Participants compare linear and nonlinear boundaries, revisit classification versus regression, and connect model outputs to loss functions, probabilities, thresholds, and evaluation metrics. A simple baseline model is used to establish a reference point before introducing a neural network.

The central case study focuses on predicting customer retention from structured behavioural data. Participants prepare numeric and categorical features, define train-validation-test splits, address class imbalance, and compare a baseline classifier with a multilayer perceptron. The emphasis is on understanding what the additional model complexity buys, rather than assuming a neural network is automatically better.

In the hands-on lab, participants build a small neural network in Keras, select activations and an appropriate loss function, inspect learning curves, and evaluate predictions using confusion matrices, precision, recall, F1-score, and ROC-AUC. They also test how changing the decision threshold affects operational outcomes.

Morning

Keynote Lecture: From Decision Boundaries to Neural Prediction

Topics covered through the keynote:

• Supervised learning, targets, and predictive tasks
• Linear versus nonlinear decision boundaries
• Classification probabilities, thresholds, and loss functions
• Train, validation, and test splits
• Baseline models and why they matter
• Feed-forward neural networks and multilayer perceptrons
• Overfitting, regularisation, and model complexity
• Evaluation beyond accuracy: precision, recall, F1, ROC-AUC

Hands-on Lab: Predicting Customer Retention

• Prepare structured data and define a predictive target
• Train a simple baseline classifier
• Build and train a neural network with Keras
• Compare learning curves and predictive metrics
• Explore the impact of changing the decision threshold
  
  

Afternoon

Group Case Study: Choosing the Right Predictive Decision

Participants work in groups with different operational costs for prediction errors and define:

• the preferred model and why
• the evaluation metric that best reflects the application
• the classification threshold and trade-offs
• evidence of overfitting or generalisation
• limitations and information needed before deployment
  
  

Group Presentations and Model Review

Short presentations (approximately 5 minutes per group)
Discussion of model complexity, reliability, and responsible use

By the end of the course, participants will be able to:

• explain the role of decision boundaries in supervised learning
• distinguish baseline predictive models from neural approaches
• prepare structured data for neural-network classification
• build and train a small feed-forward neural network in Keras
• evaluate classifiers using multiple predictive metrics
• reason about probability thresholds and error trade-offs
• identify signs of overfitting and justify an appropriate model for a given use case
  
  

One-day intensive workshop combining:

• conceptual keynote lecture
• guided predictive-modelling exercises in Jupyter notebooks
• comparison of classical and neural models
• group-based decision and threshold challenge
• presentations and critical discussion