The AI Pipeline and Learning Paradigms

The AI Pipeline: From Data to Decision

Every AI product runs the same pipeline — understanding it exposes where things break and what vendor claims to probe:

• Collect & clean → raw data is gathered and prepared; quality here determines quality everywhere downstream
• Train → a model learns patterns from the data and its weights are frozen
• Deploy → the frozen model serves real-time inference; the world keeps changing, the model does not
• Feedback loop → outcomes flow back into future retraining — without this, the model is a depreciating asset
• Key question for any vendor: is there a feedback loop, or is this a snapshot that never improves from production signals?

Three Ways AI Learns

The learning paradigm determines what data the model needs, what it is good at, and where it fails. All three are in production today.

• Supervised: trains on labelled examples ("this is X, this is not X"). Most classical ML works this way — spam filters, classifiers, scoring models. Accuracy depends entirely on label quality.
• Unsupervised: finds patterns without labels — clustering, anomaly detection, structure discovery. Strength: catches unknown unknowns. Weakness: high false-positive rates without careful tuning.
• Reinforcement: learns from reward signals as it tries actions. Used in agentic systems and LLM post-training (RLHF). This is how chatbots learn to be helpful and game AIs surpass humans.

Neural Networks: Layers of Weighted Transforms

A neural network is layers of mathematical neurons. Each applies a weighted transform plus a non-linear function. Stack enough layers and the network can approximate extremely complex functions — that is the whole trick. Training works through three steps: forward pass (data flows through, prediction comes out), loss function (measures how wrong the prediction was), and backpropagation (pushes corrections backwards through the network, adjusting weights). Repeat millions of times. The trained model is just the final set of weights — a frozen snapshot of patterns learned from data. It does not "know" things in a human sense; it computes statistically likely outputs.

The Transformer: Why Everything Changed in 2017

Deep learning applies across three domains — Natural Language Processing (text), Computer Vision (images and video), and Robotics/Automation (perception and control). All three were transformed by a single 2017 architecture: the Transformer.

• NLP: text understanding and generation — chatbots, translation, summarisation, and the foundation of LLMs
• Computer Vision: image recognition and object detection — visual classification, screenshot analysis, CCTV analytics
• Robotics and Automation: perception, navigation, and control systems
• The Transformer's attention mechanism allowed models to learn context across long sequences — enabling the leap from narrow task models to broad-domain LLMs
• Everything since 2022 (ChatGPT, Claude, Gemini) is built on this architectural foundation