B4 · Domains of AI: Vision, Language, Speech and Robotics

The Three Core Domains: Data, Vision and Language

AI divides into fields according to the kind of problem. Natural language processing handles text, computer vision handles images, speech AI handles audio, robotics acts in the physical world, recommender systems rank items, and planning and search choose sequences of actions.

10 min Beginner Jump to the activity

What you need: Nothing. This page starts from the beginning, with no programming and no formulas.

Definition

AI divides into fields according to the kind of problem. Natural language processing handles text, computer vision handles images, speech AI handles audio, robotics acts in the physical world, recommender systems rank items, and planning and search choose sequences of actions.

In simple words

Think of a human: reading is NLP, seeing is vision, hearing is speech, moving is robotics, and deciding what to do next is planning. AI has a field for each.

Why it matters

It's your map of this course. Most jobs and projects sit in one or two fields, so you can choose where to go deep.

Key terms

Domain
A branch of AI defined by the kind of data it works with.
Data science
Finding patterns in numbers and tables in order to predict, recommend and decide.
Computer vision
The domain of AI that works with images and video.
Natural language processing
The domain of AI that works with human language, written or spoken.
Modality
One form of data: numbers, text, image, audio or video.
Multimodal AI
A system that can take in or produce more than one form of data, such as a photo and a question about it.

Explained in detail

Why AI is divided into domains

AI is a wide field, and its problems differ mainly in the kind of data they involve. A table of sales figures, a photograph and a sentence look nothing alike to a computer, and for many years each needed its own methods, its own experts and its own research community. That is how AI came to be divided into domains. The division is by input: what kind of data goes in decides which domain a task belongs to.

The three core domains

DomainInputTypical questionsEveryday example
DataNumbers and categories in tablesHow much? Which group? What next?Price predictions, fraud checks, recommendations
Computer visionImages and videoWhat is in this picture, and where?Face unlock, scanning a document, reading a number plate
Natural language processingText and speechWhat does this mean? What should the reply be?Translation, search, chatbots, voice typing

The data domain is the oldest and the most widely used, although it is the least visible. Whenever a system looks at rows of past records and estimates a number or a category, it belongs here. Computer vision gives machines a sense of sight, and natural language processing lets them work with the words people use. Speech is usually counted with language, because it deals with the same words in spoken form.

One idea underneath all three

Although the inputs look different, the three domains now share the same core. In each, the input is turned into numbers: a table already is numbers, an image becomes pixel values, and text becomes numbered tokens. A model is trained on many examples, finds patterns in those numbers and produces an output. This is why progress in one domain often spreads to the others. The network design invented for translating language in 2017 was soon used for images, sound and protein structures as well.

Domains that build on these

  • Speech: turning voice into text and text into voice.
  • Robotics: using vision, data and sometimes language to act in the physical world.
  • Generative AI: creating new text, images, audio and video. It cuts across all the domains.

The walls are coming down

For a long time a vision system could not read and a language system could not see. Today's largest models are multimodal: you can show one a photo of a broken tap and ask, in words, how to fix it, and it answers using both. This does not make the domains useless. They remain the best way to describe what a system is doing at each step, and to find where a mistake came from. When an app misreads a signboard and then translates the wrong word perfectly, it helps to know that the failure was in vision and not in language.

How it works

FieldInputExampleModule
NLPTextTranslation, sentimentNatural Language Processing
Computer visionImages, videoDetection, OCRComputer Vision
SpeechAudioVoice typing, text-to-speechSpeech and Audio AI
RoboticsSensorsWarehouse robotsRobotics
RecommendersBehaviourProduct suggestionsApplied and Advanced ML
Planning, searchGoals, statesRoutes, schedulesAgents and Search
Generative AIPromptsWriting, imagesLarge Language Models

Try it yourself

AI's fields, like a person's senses

A home robot makes you tea. Each ability it uses is its own field of AI.

Which domain is it?

Look at what each system takes in: numbers and records, pictures, or human language.

  • Suggesting a film from what you've watched

  • Predicting tomorrow's sales from past sales

  • Deciding whether a review is positive

  • Reading the text from a scanned page

  • Translating Hindi into English

  • Finding the lanes on a road from a car camera

Common mistakes

  • Trying to learn every field at once. Learn the common base, then choose one.
  • Thinking these are separate worlds. They share maths and methods.

Check yourself

Question 1

1. Reading text from a scanned page falls under:

Question 2

2. Suggesting a film falls under:

Summary

  • Fields follow the type of problem and input.
  • Learn the shared base first, then pick a field.
  • Multimodal models are blurring the borders.

Finished reading?