By ShilpBot Team · 7 min read · Updated
Artificial intelligence, or AI, is the field of building software that performs tasks which normally need human judgement, such as recognising speech, translating languages, or deciding which route to take. The term covers a wide range of techniques, from hand-written rules to systems that learn from data.
AI, machine learning and deep learning
These three terms are nested, like circles inside circles.
| Term | Meaning | Example |
|---|---|---|
| Artificial intelligence | Any technique that lets machines show intelligent behaviour | A chess program using search rules |
| Machine learning | AI that learns patterns from data instead of following only fixed rules | A spam filter trained on labelled emails |
| Deep learning | Machine learning using many-layered neural networks | Image recognition in a phone camera |
Rules versus learning
Early AI systems relied on rules written by experts: if the temperature is above a threshold, then do something. This works for narrow, well-understood problems. It breaks down when a task is too messy to describe in rules. Nobody can write a complete rule set for recognising a cat in a photo.
Machine learning flips the approach. Instead of writing rules, you give the system many examples and let it find the pattern itself.
Three main ways machines learn
Supervised learning
The model learns from examples that come with the correct answer, called labels. Show it thousands of photos tagged "cat" or "dog" and it learns to tell them apart. Most practical AI today, including image classification and price prediction, uses supervised learning.
Unsupervised learning
There are no labels. The model looks for structure by itself, such as grouping customers with similar behaviour. It is useful for exploring data when you do not yet know what to look for.
Reinforcement learning
An agent learns by trial and error, receiving rewards for good outcomes and penalties for bad ones. This approach is common in game-playing systems and in teaching robots skills such as walking or grasping in simulation.
How a model is built, step by step
- Collect data relevant to the problem.
- Prepare the data by cleaning mistakes and splitting it into training and testing sets.
- Train the model so it adjusts its internal numbers to reduce its mistakes on the training set.
- Evaluate it on the test set, which it has not seen, to check how well it generalises.
- Deploy and monitor it, since real-world data changes over time.
What AI does well, and where it struggles
AI is strong at pattern recognition at scale, repetitive classification and prediction when plenty of good data exists. It is weaker when data is scarce or biased, when situations differ from its training data, or when it needs to explain its reasoning.
Limits worth knowing
- Bias: a model trained on skewed data produces skewed results.
- Overfitting: the model memorises its training data and performs badly on new data.
- Confident errors: some models, including language models, can produce fluent but wrong answers. Important outputs should be verified by a person.
Why this matters for robotics
Robots use AI mostly for perception, such as recognising objects and understanding scenes, and for decision-making, such as planning where to move. Learning how models work gives you the foundation for the lessons in AI for Robotics.
What to learn next
To see what happens inside the most popular kind of model, read How Neural Networks Learn.
Frequently asked questions
Is machine learning the same as artificial intelligence?
No. Machine learning is one approach within the broader field of AI. AI also includes rule-based systems, search and planning methods that do not learn from data.
Do I need advanced maths to start learning AI?
Not to understand the concepts. Comfortable use of basic algebra and some familiarity with probability helps, and deeper study later involves linear algebra and calculus, but you can build a solid intuition first.
What is the difference between training and testing data?
Training data is what the model learns from. Testing data is held back and used only to measure how well the finished model handles examples it has never seen.
