AI Search 101
What is deep learning?
Deep learning is machine learning using neural networks with many layers. The depth lets the system build its own representation of the data, learning simple features in early layers and combining them into complex ones later, which removes the need for a person to specify what to look for.
The basics
| Part of | Machine learning, itself part of artificial intelligence |
| Built from | Neural networks with many hidden layers |
| Key advantage | Learns its own features rather than being told what matters |
| Requires | Large data volumes and substantial computing power |
| Behind | Large language models, image generation, speech recognition |
What does the depth actually buy?
Feature learning, and it is the whole point.
Before deep learning, applying machine learning to images meant a person deciding what the system should measure: edges, corners, colour histograms, textures. This was called feature engineering, it required domain expertise, and the ceiling on performance was the quality of the human's ideas.
A deep network works it out instead. Early layers learn to detect simple things such as edges and gradients. Middle layers combine those into shapes and parts. Later layers combine parts into objects. Nobody specifies this hierarchy; it emerges because it is an efficient way to solve the task.
The same principle applies to language. Early layers capture something like syntax and word form, later layers capture meaning, reference and the relationships between distant parts of a passage.
Why does deep learning need so much data?
Because it is learning the representation as well as the task, and representations are expensive.
A model with a handful of hand-designed features has few parameters and can be fitted from modest data. A deep network with billions of parameters has enormous capacity, and capacity without sufficient data produces memorisation rather than generalisation. The volume of examples is what forces it to find patterns that actually hold.
This is also why transfer learning matters commercially. A model trained once on an enormous general corpus can be adapted to a specific task with comparatively little data, because the expensive representation-learning has already happened. Almost every practical application today starts from a pre-trained model rather than from scratch.
Where does deep learning fail?
Small data. With a few thousand rows of tabular data, simpler methods usually match or beat it and are easier to reason about.
Explanation. The learned representation is distributed across the whole network, so there is no clean account of why a particular decision was made. In regulated settings this is a genuine constraint rather than an inconvenience.
Out-of-distribution input. Performance degrades on data unlike anything in training, and it degrades without warning. The system does not know that it is now guessing.
Cost. Training large models is expensive enough to concentrate the capability among a small number of organisations, which is itself a strategic fact worth noticing.
How deep learning is used in marketing
Production. Every generative tool in use rests on deep learning, from text and image generation to voice synthesis and translation.
Analysis. Sentiment and intent classification on open text, image and video recognition, speech transcription, and clustering unstructured feedback.
Distribution. The recommendation and ranking systems inside advertising and social platforms, which decide reach far more than any manual targeting setting does.
How this impacts your business
A language model does not look a company up in a register. It composes a description from everything it read during training and whatever it retrieves at the moment of the question. Trade press, directories, forums, competitor comparison pages, and somewhere in that mix the company's own website too.
So the description a buyer sees is assembled, not published. No company controls it, but every company supplies a share of the material it is built from.
The traffic impact is still small, roughly one percent of sessions for most companies. The influence is not small, because the reader is often a buyer forming a first impression of what a company does and who it serves, before any page gets a click.
Checking takes ten seconds. Type the company name into ChatGPT and read the answer.
Frequently asked questions
How many layers make a network deep?
There is no formal threshold. Historically more than two or three hidden layers counted as deep; current models have dozens to hundreds.
Is deep learning always better than other machine learning?
No. It dominates on unstructured data such as text, images and audio. On modest structured datasets, gradient boosted trees and similar methods frequently perform better and are cheaper to run.
What is transfer learning?
Starting from a model trained on a large general dataset and adapting it to a specific task. It is how most applied deep learning is done, because it removes the need for both the data and the compute of full training.
Does deep learning require labelled data?
Not always. Large language models are trained largely by predicting missing or subsequent text, which needs no labels, and are then refined with smaller labelled sets.
Why is deep learning called a black box?
Because the knowledge is spread across billions of parameters with no human-readable structure. The behaviour can be characterised statistically, but individual decisions cannot be traced the way rule-based logic can.
Related definitions
- What is a neural network?
- What is machine learning?
- What is a large language model?
- What is generative AI?
- AI Search 101
Written by Lari Numminen, Generate More. Also available in Finnish. Updated 31 August 2026.