Learn
Artificial
Intelligence

Tables, text, images, audio. All of it is data. Show how AI sees the world purely as numbers. Introduce real datasets: CSV files, images as pixel arrays, text as tokens. Students pick apart a sample dataset together live.

Mean, median, variance, distributions, explained visually. Why does it matter? Because AI models are sophisticated statistics engines at their core. Real datasets, histograms and scatter plots. No proofs, no memorisation.

Introduce linear algebra as the language AI speaks. A vector is just a list of numbers. A matrix is a table. Kept visual with arrows and grids. Demonstrate how images are simply matrices of pixel values.

Set up Google Colab with zero installation. Learn variables, lists, loops, functions. Load a CSV with Pandas. Plot data with Matplotlib. First hands-on code session. Keep it celebratory, not stressful.

Work with a real messy dataset. Handle missing values, duplicates, outliers. Visualise distributions. Students learn that 80% of real AI work is data prep, not modelling. Far better to know this upfront.

Introduce the learning loop: data → model → prediction → error → adjust. Use a simple linear regression example with a live visual of the line "fitting" to points. The session where everything clicks for beginners.

Spam vs not spam, cat vs dog. Decision trees, K-Nearest Neighbours explained visually. Build a classifier on the Titanic dataset. Will this passenger survive? Students see and own their first real model prediction.

Predict house prices with linear and polynomial regression. Plot the prediction line live. Introduce loss functions and ask: how do we measure exactly how wrong a model is? MSE explained visually with no derivations.

No labels, no right answers. K-Means clustering: let the model find groups entirely on its own. Customer segmentation as the example. Visualise clusters with colour-coded scatter plots. Students discover patterns they didn't put there.

Why you always split your data. Train/test/validation sets. The bias-variance tradeoff. Overfitting illustrated clearly: a model that memorises vs one that generalises. Cross-validation basics with Scikit-learn in practice.

Accuracy alone is not enough. Precision, recall, F1-score, confusion matrices explained. When does a false negative cost more than a false positive? Cancer detection vs spam filter. Context completely changes what "good" means.

Start with a biological neuron analogy students already understand. Show a perceptron: inputs, weights, bias, activation function. Use TensorFlow Playground so students can visually tune a single neuron and watch it learn in real time.

Stack neurons into layers: input, hidden, output. Why depth matters so much. Introduce activation functions: ReLU, Sigmoid and Softmax, with guidance on when to use each. Build a simple network in Keras in under 10 lines of code.

How filters scan an image and build up feature maps. Max pooling, local connectivity, and weight sharing. CNN architecture from input to classification output. Landmark models from LeNet-5 to ResNet. Build an MNIST digit classifier in Keras achieving 99.2% accuracy.

Self-attention explained from first principles: queries, keys, and values. Scaled dot-product attention formula and why it works. Multi-head attention, positional encoding, and the full encoder block. BERT vs GPT. Use Hugging Face pipelines for sentiment analysis and text generation in 5 lines of code.

Why training from scratch is almost never the right choice. Feature extraction vs fine-tuning, when to use each. Fine-tune ResNet-50 on a custom image dataset in two phases with Keras. Fine-tune DistilBERT for text classification with Hugging Face Trainer. Understand catastrophic forgetting and how to avoid it.

Where bias originates in the AI pipeline, and why it compounds. Six bias types with real case studies: COMPAS recidivism, Amazon's hiring tool, Obermeyer's healthcare disparity (Science 2019), and Gender Shades. The mathematical impossibility of satisfying all fairness definitions simultaneously. Explainability tools. EU AI Act and regulation.

Tokenisation and Byte-Pair Encoding. Next-token prediction as the core training objective. Scaling laws (Kaplan 2020, Chinchilla 2022) and emergent abilities. RLHF, Constitutional AI, and DPO alignment techniques. Hallucination types and mitigation. RAG architecture. Context windows from 4K to 1M tokens.

Zero-shot, few-shot, and chain-of-thought prompting with verified research citations. System prompts and role prompting. Temperature, top-p, top-k, and when to adjust them. Practical techniques, common mistakes, and full working OpenAI API v1.0+ code including structured JSON output with chain-of-thought.

Classification, detection, segmentation, and pose estimation. YOLO's single-pass detection architecture and its evolution through v11. U-Net for medical image segmentation. OpenCV for image pre-processing and result overlay. Hugging Face DETR and torchvision ResNet-50 inference pipelines. Medical imaging challenges, DICOM files, and ethical considerations in facial recognition.

Multimodal AI (CLIP, GPT-4o, diffusion models). AI agents, the ReAct loop, and function calling with full Python code. Reasoning models: OpenAI o1, o3, and DeepSeek-R1. Open vs closed model trade-offs. AI safety, alignment, and mechanistic interpretability. Economic impact studies. Where to go next as a practitioner.