What Is Deep Learning?

Yann LeCun, Yoshua Bengio and Geoffrey Hinton summarised the field in Nature in May 2015. Their starting point is a limit of the older methods: conventional machine-learning techniques were limited in their ability to process natural data in raw form. For decades, building a recognition system took careful engineering and considerable domain expertise to design a feature extractor, the component that turns, for instance, the pixel values of an image into something the model can classify.

Deep learning removes that manual step. The authors explain it with an image. The first layer learns to detect edges at particular orientations and locations. The second spots motifs, particular arrangements of edges. The third assembles motifs into parts of familiar objects, and later layers detect objects as combinations of those parts. The key aspect, they write, is that these layers of features are not designed by human engineers: they are learned from data using a general-purpose learning procedure.

"Deep" refers to the number of layers. Google's machine learning glossary calls a neural network with more than one hidden layer a deep model. How artificial intelligence, machine learning and deep learning relate to one another is covered in the entries on machine learning and on the neural network: deep learning is the part of machine learning that works with deep neural networks.

A deep model has weights, numbers that set how strongly each connection between layers counts. According to LeCun, Bengio and Hinton, a typical deep-learning system may have hundreds of millions of these adjustable weights and hundreds of millions of labelled examples to train it with. At each training step, the algorithm works out, for every weight, how the error would change if that weight rose slightly, and moves it the other way. The calculation runs backwards layer by layer, using the backpropagation method described in the neural network entry.

That gives deep learning its two requirements: data and computing power. A model that learns its own features needs many examples so that it does not learn features that only hold for a few cases. And adjusting hundreds of millions of weights takes a great many operations. The authors write that speech recognition was the first major application of the approach, made possible by graphics processing units (GPUs) that were convenient to program and let researchers train networks 10 or 20 times faster. They expected further progress precisely because deep learning needs very little engineering by hand and can easily take advantage of more computation and data.

One well-known case shows the scale. In 2012 Alex Krizhevsky, Ilya Sutskever and Geoffrey Hinton trained a deep convolutional network on the 1.2 million images of the ImageNet contest from 2010, spread across 1,000 classes. The network had 60 million parameters and took five to six days to train on two GPUs. The authors wrote that their results could be improved simply by waiting for faster GPUs and bigger datasets. In the 2015 Nature review, LeCun, Bengio and Hinton report that such networks almost halved the error rates of the best competing approaches in the 2012 ImageNet competition.

According to the same review, deep learning has dramatically improved speech recognition, visual object recognition and object detection. On 29 September 2021 Google announced that MUM, a model that understands text and images, would let people search in Google Lens by combining an image and a written question. LLMs are deep models too, trained on large volumes of text, and belong to natural language processing.

Google Search has used deep learning since 2015. Pandu Nayak wrote on Google's blog on 3 February 2022 that RankBrain, launched in 2015, was the first deep learning system deployed in Search. On 25 October 2019 he had announced BERT, a neural network-based technique that came out of Google's research on transformers, models that process each word in relation to all the other words in a sentence. Google said BERT would help Search better understand one in ten searches in the US in English. On 18 May 2021 he introduced MUM and called it 1,000 times more powerful than BERT. Google's guide to its ranking systems, last updated on 10 December 2025, states that MUM is currently used for some specific applications rather than for general ranking.

For SEO, the practical consequence is that Google interprets what a query means and what a page answers. The guide describes BERT as a system for understanding how combinations of words express different meanings and intent, and describes no BERT-specific optimisation. What does change is the kind of text worth writing: one that answers the question, not one that repeats the keyword in every variant.

There is also a limit worth knowing. Google's glossary notes that most deep models and large language models are black boxes: you can see how the answer changes with the input, but you cannot work out exactly how the model arrives at it. No one outside Google can deduce which feature led to a particular ranking.

• Only call a system deep learning if it uses deep neural networks, not any AI system.
• Before using a deep learning tool, check which data it was trained on and whether it resembles your own.
• Assess its output against a sample of your own, reviewed by people.
• Write content that answers the user's question instead of stacking keyword variants.
• Cite model figures (parameters, data, improvements) with their source and date.

• Using deep learning as a synonym for artificial intelligence.
• Assuming more layers always give better results, without more data or more computing power.
• Looking for an "optimisation for BERT" or MUM that Google does not describe.
• Reading a deep model's answer as an explanation of why it gave that answer.
• Assuming MUM decides the order of search results.