AI, Machine Learning and Deep Learning
AI, machine learning and deep learning are related, but they are not interchangeable. Think of them as nested layers: deep learning is one approach within machine learning, and machine learning is one approach within the wider field of AI.
Learning objectives
What you will be able to do
- Explain ai, machine learning and deep learning accurately in your own words.
- Recognise how the concept appears in real AI products and professional work.
- Identify an important limitation, risk or evaluation requirement.
- Apply the concept through a practical activity and knowledge check.
Core explanation
Build the right mental model
Artificial intelligence is the broad goal of creating systems that perform intelligent tasks. Some AI systems use hand-written rules. Machine learning instead allows a system to learn useful patterns from data rather than programming every decision explicitly.
Deep learning uses neural networks with many processing layers. It became especially important for images, speech, language and other problems where useful features are difficult to specify manually. Large language models are deep-learning systems trained to predict and generate sequences of tokens.
The best method is not automatically the newest one. A clear rule or simple statistical model may be cheaper, more transparent and more reliable than a large neural network.
Classify the mechanism, not the marketing label
AI names a broad field. Machine learning is a family of methods within it that fit patterns from examples. Deep learning is a family of machine-learning methods based on neural networks with multiple layers. These relationships help you ask sensible technical questions, but they do not tell you whether a particular system is useful. A simple approach can solve a task that does not need a complex one.
Consider three versions of a document-routing tool. Version A follows expert-authored rules about terms in the subject line. Version B learns a category mapping from labelled examples. Version C uses a multilayer network to process text representations. B is machine learning; C is both machine learning and deep learning. A could be part of a symbolic AI system, although an ordinary keyword rule alone is not necessarily a meaningful AI product.
Training and use are different stages
A learned model has a training stage in which parameters are adjusted and a use stage, often called inference, in which it produces outputs from inputs. Sending a new document through a fixed model does not automatically retrain it. An application may save feedback for a later training process, but that is an additional mechanism that must be established rather than assumed.
Generative AI describes the kind of output being produced, such as new text or images. It is not a separate circle that cleanly replaces the AI–ML–deep-learning relationship. A text-generating neural model can belong to all three categories. Likewise, an application may combine rules, retrieval, a learned classifier and a text generator in a single workflow.
Compare candidates against the same task
Write a baseline before evaluating a complex model. A baseline might be a transparent rule, a constant prediction or the current manual workflow. Measure the same outcomes for all candidates using the same evaluation cases. Otherwise, a comparison can reward a system merely because it received easier examples or a different definition of success.
Include operation costs in your comparison: response time, maintenance, data access, ease of correction and the consequence of errors. A model’s sophistication is not a substitute for these checks. If the organisation changes a routing policy tomorrow, a rule may be easy to update; a learned approach may need new examples, evaluation and deployment. In other tasks, manually maintaining rules may become impractical.
AI
The broad discipline and objective: machines performing tasks associated with intelligence.
Machine learning
Algorithms improve task performance by finding patterns in data.
Deep learning
Machine learning based on multi-layer neural networks.
Generative AI
Models that create new text, images, audio, video, code or other content.
A fictional comparison with a useful baseline
Suppose a team evaluates three routing candidates on the same 40 synthetic requests. The rule system is correct on 30, a simple learned classifier on 34 and a deep network on 35. Their respective accuracies are 75%, 85% and 87.5%. The deep model gets one additional case correct compared with the simple classifier, not a dramatic transformation supported by the architecture name alone.
Now inspect which case improved and which cases worsened. If the extra correct case is routine but the deep system makes a more disruptive error elsewhere, the aggregate score may conceal an important trade-off. The team should also ask whether 40 cases are enough to represent future requests. These fictional figures demonstrate comparison, not measured evidence that one model family is superior.
Apply your learning
Practical activity
- In five minutes, describe a task with a clear rule-based baseline.
- In five minutes, propose one learned alternative and identify the examples it would need.
- In five minutes, make a comparison table containing correctness, review workload, response time and maintenance effort.
- In five minutes, state what evidence would justify moving from the baseline to the learned alternative.
Compare with an explained example answer
For a fictional equipment-request queue, a baseline can route requests by an explicit equipment category selected by the requester. A learned alternative could infer that category from free text, using approved labelled request descriptions.
Keep the baseline when the form is reliable and free-text inference adds little. Consider the learned alternative when many requests lack a category and evaluation shows a useful reduction in manual routing without unacceptable errors. The decision depends on evidence rather than whether a vendor uses the phrase deep learning.
Write your answers in a learning journal. The value comes from connecting the concept to your own profession.
Knowledge check
Test your understanding
Q1Is an LLM outside machine learning because it generates text?
Answer: No. Generation describes its output; an LLM is normally a deep-learning model and therefore also a machine-learning model.
Q2Does using a model necessarily change its parameters?
Answer: No. Ordinary inference can run a fixed model. Training or adaptation requires an additional process.
Q3Why compare a model with a simple baseline?
Answer: It shows whether the added complexity produces enough benefit for the actual task and operating constraints.
If you cannot explain the answer without reading it, revisit the core explanation and example before continuing.
Continue with authoritative sources
Important industry resources
These links lead to official organisations, industry laboratories or established open-source learning projects.