2.3 Types of AI Technologies

2.3 Types of AI Technologies

Artificial intelligence is a family of techniques, not a single product feature. Two systems can both be called AI and still fail in opposite ways: one clusters customers with no labels, another plays a game by maximizing a reward, a third fires a 1980s expert-system rule. If you cannot name the technology, you cannot name the oracle, the data risk, or the retraining story.

This section is a tester's map. You are not asked to derive back-propagation on the exam. You are asked to explain the types, recognize what each is for, and notice how modern systems stack them.

Machine learning as a core branch

Machine learning lets a system learn from data and build a model without a person writing a complete conventional rule list for the decision. The learning recipe still has plenty of human choices—features, architecture, loss, stopping rules—but the detailed mapping from input to output is estimated from examples.

Not every ML system keeps learning forever. Some machine learning systems stay locked after training: they serve the knowledge they acquired and need an explicit retrain to change. Others adapt continuously (or on a frequent schedule) as new data arrives. That contrast is a preview of locked versus adaptive systems you will test later. Ask it on day one of a project: after go-live, who is allowed to change the model's parameters, how often, and how will we notice a bad change?

ML itself splits into three approaches you must be able to tell apart.

Supervised learning

Supervised learning uses labeled data. Each training row (or image, or snippet) comes with a target the system should imitate: spam or not, house price, diagnosis code, remaining useful life.

Classic algorithms include linear regression (predict a number along a fitted relationship) and decision trees (learn a flowchart of splits from the labels). Modern neural nets can be supervised too. Typical jobs are classification (choose a class) and prediction of a quantity or a risk score.

Tester notes:

• Label quality is part of the product. Wrong labels train the wrong policy.
• You need held-out data that the team did not tune against, or you will celebrate memorization.
• Class imbalance (1% fraud) makes accuracy a vanity metric; you will meet better metrics later in this guide.

Unsupervised learning

Unsupervised learning uses unlabeled data. There is no official y column. The algorithm looks for structure: groups, unusual points, compressed representations.

Clustering is the example to remember: group customers, log lines, or cells that look alike. The oracle is weaker than in supervised work. There is often no single correct cluster count. Testers lean on stability (do clusters jump when you resample?), usefulness to a downstream task, and sanity checks from domain experts.

Reinforcement learning

Reinforcement learning (RL) trains an agent that takes actions in an environment and receives rewards and penalties. The agent is not primarily mimicking a labeled spreadsheet. It is learning a policy that accumulates reward: a robot that docks a charger, a bidding agent, a game player, a traffic-light controller.

Tester notes:

• The reward function is a specification. If you reward speed only, you will get reckless speed.
• Exploration can be unsafe on live machinery; teams often train in simulation first.
• Non-determinism is the default: the same state can yield different actions while the policy is still learning, and even a frozen policy may be stochastic.

Key ML technologies you will see on architecture diagrams

Beyond the three learning setups, several model families appear again and again:

• Neural networks — layered numerical transformers of features; the workhorse behind much modern perception and language work
• Bayesian models — reason with probabilities and prior beliefs; useful when you need calibrated uncertainty or small-data caution
• Support vector machines (SVM) — find decision boundaries with a well-studied geometric recipe; still common on medium tabular and text-feature problems
• Random forests — ensembles of trees that vote; strong baselines on tabular data and less of a black box than a giant net

Testers should not worship a family. A random forest that is well validated can beat a fashionable network that was trained on leakage. Ask what input it needs, how it fails, whether it outputs a probability you can threshold, and how you will retrain it.

Deep learning as a subset

Deep learning (DL) is a subset of ML that uses deep neural networks—many layers—to attack problems where raw signals (pixels, waveforms, tokens) are too messy for hand-built features. Three architectures are the vocabulary the exam expects:

• Convolutional neural networks (CNN) are highly effective for image recognition and object detection. Convolutions slide local filters across the image, which matches the way nearby pixels relate. Testers still worry about lighting, occlusion, camera change, and tiny adversarial stickers.
• Recurrent neural networks (RNN) specialize in sequential data such as text or time series. They carry a hidden state forward through the sequence. They can struggle with very long-range dependencies, which is one reason transformers took over much of language work.
• Transformers handle long-range dependencies in sequences using self-attention. They power a huge fraction of modern natural language processing and, as vision transformers, are used on images as well.

If a stakeholder says we use AI, ask which of these (or which older method) is actually in the path. A CNN that only classifies a bolt as cracked is a different test object from a transformer that writes a narrative about the bolt.

Other AI technologies that are not just deep nets

The field did not begin with transformers, and products still mix older ideas:

• Natural language processing (NLP) analyzes language: sentiment, translation, entity extraction, topic labels. NLP can be classical (dictionaries, statistical n-grams) or deep (transformers). Either way, language shift and dialect coverage are test data problems.
• Computer vision analyzes visual data for applications such as facial recognition, inspection, and robotics. Vision is a task area; CNNs and vision transformers are tools inside it.
• Fuzzy logic reasons under uncertainty with degrees of truth rather than strict true/false cuts. It still appears in control systems where engineers want inspectable graded rules.
• Search and optimization algorithms explore large action spaces: route planning, scheduling, strategic decision search. Failures look like timeouts, locally awful plans, or constraint violations, not like a wrong class label.
• Rule-based reasoning / expert systems encode specialist knowledge as structured rules for decision support. They are AI in the historical sense and conventional-looking in the testing sense: you can often trace the fired rules. They are not deep learning, even if a vendor slide says AI.

Full integration of all of these into one mind is still limited. What you will see is composition: an expert-system override on top of a model score, a search planner that calls a vision classifier, an NLP component that writes a ticket the workflow engine routes.

Generative AI, large language models, and agentic AI

Generative AI builds on the technologies above to create new content—text, images, audio, and more (the next section goes deep). Large language models (LLMs) combine deep neural networks with NLP to analyze and generate human-like language. They are not a third physics. They are a particular stacking of deep nets and language objectives, usually transformer-based.

Agentic AI extends these pieces through autonomous agents that plan, reason, and act to pursue goals in changing environments. An agent might break a user goal into steps, call tools (search, calendar, code execution), observe results, and continue. For testers, agency multiplies side effects: the defect is not only a wrong sentence; it is a wrong purchase, a wrong email, a wrong configuration change. You test permissions, stop conditions, and traces, not only BLEU scores or vibe.

flowchart TB
  AI[Artificial intelligence]
  ML[Machine learning]
  DL[Deep learning]
  Other[NLP vision fuzzy search expert systems]
  Gen[Generative AI and LLMs]
  Agent[Agentic AI]
  AI --> ML
  AI --> Other
  ML --> DL
  DL --> Gen
  Other --> Gen
  Gen --> Agent
  ML --> Agent

Putting the map to work on one product

Imagine a warehouse robot. Computer vision (often a CNN) sees a tote. A supervised model classifies SKU and damage. Unsupervised clustering might later discover a new damage pattern nobody labeled. A planner (search/optimization) chooses a path. Fuzzy or rule-based logic may slow the robot in a human-shared aisle. An LLM generates an incident summary. An agent might request a replenishment order. Some of those models are locked between releases; the path planner might re-optimize every second.

Your test plan should name each technology, its input contract, whether it can change at runtime, and what oracle is honest. If you write one generic AI test case for that robot, you have not started.