McKinsey Explainers

What is deep learning?

Deep learning is a type of machine learning that is more capable,
autonomous, and accurate than traditional machine learning.

April 2024
Deep learning has been around for a while, but instruction. These algorithms can detect patterns
most of us never used a deep learning–based tool and learn how to make predictions and
until the release of OpenAI’s ChatGPT, in late 2022. recommendations by processing data and
(And even as we marveled at ChatGPT’s outputs, experiences, rather than by receiving explicit
most of us didn’t know it was using deep learning to programming instruction. The algorithms also adapt
generate them.) Like its predecessors DALL-E, in response to new data and experiences to
Google’s Imagen and PaLM, Stable Diffusion, and improve over time.
others, ChatGPT relies on large deep learning
models trained on massive data sets to generate The volume and complexity of the data that is now
content based on prompts. But unlike its being generated, too vast for humans to reckon
predecessors, ChatGPT works via an open-access with, has increased the need for machine
API, which means the general public can learning—and has enhanced its potential. In the
experience the power of deep learning for the years since its widespread deployment, machine
first time. learning has had impact in a number of industries,
including medical-imaging analysis and high-
The world of artificial intelligence and machine resolution weather forecasting.
learning (for which deep learning is the next
evolutionary step) is undergoing a generational For more on machine learning, check out our
transformation, from an idea studied by scientists to McKinsey Explainer.
a tool used by all kinds of people for all kinds of
tasks. McKinsey analysis has shown that between
2015 and 2021, the cost to train an image How is deep learning different from
classification system (which runs on deep learning machine learning?
models) fell by 64 percent. Training times improved Deep learning is a more advanced version of
by 94 percent in the same period. We’ve also machine learning that is particularly adept at
found that generative AI (gen AI) could add the processing a wider range of data resources (text, as
equivalent of up to $4.4 trillion annually to the well as unstructured data including images),
global economy. These profound changes are all requires even less human intervention, and can
powered by deep learning. often produce more accurate results than
traditional machine learning. Deep learning uses
But what actually is deep learning? And how does it neural networks—based on the ways neurons
make all this possible? Read on to find out. interact in the human brain—to ingest and process
data through multiple neuron layers that recognize
Learn more about McKinsey Digital. increasingly complex features of the data. For
example, an early neuron layer might recognize
something as being in a specific shape; building on
What is machine learning? this knowledge, a later layer might be able to
Before we move to deep learning, let’s get the identify the shape as a stop sign. Similar to machine
basics down. Machine learning is a form of artificial learning, deep learning uses iteration to self-correct
intelligence that can adapt to a wide range of and to improve its prediction capabilities. Once it
inputs, including large data sets and human “learns” what an object looks like, it can recognize
the object in a new image.

What is deep learning? 2

What’s the relationship between deep What kinds of neural networks are
learning and gen AI? used in deep learning?
ChatGPT made AI visible—and accessible—to the There are three types of artificial neural networks
general public for the first time. ChatGPT, and other used in deep learning:
language models like it, were trained on deep
learning tools called transformer networks to — Feed-forward neural network. In this simple
generate content in response to prompts. neural network, first proposed in 1958,
Transformer networks allow gen AI tools to weigh information moves in only one direction: forward
different parts of the input sequence differently from the model’s input layer to its output layer,
when making predictions. Transformer networks, without ever traveling backward to be
comprising encoder and decoder layers, enable gen reanalyzed by the model. That means you can
AI models to learn relationships and dependencies feed, or input, data into the model, then “train”
between words in a more flexible way compared the model to predict something about different
with traditional machine and deep learning models. data sets. As just one example, feed-forward
That’s because transformer networks are trained on neural networks are used in banking, among
huge swaths of the internet (for example, all traffic other industries, to detect fraudulent financial
footage ever recorded and uploaded) instead of a transactions. Here’s how it works: first, you train
specific subset of data (certain images of a stop a model to predict whether a transaction is
sign, for instance). Foundation models, as further fraudulent based on a data set you’ve used to
discussed below, trained on transformer network manually label transactions as fraudulent or not.
architecture—like OpenAI’s ChatGPT or Google’s Then you can use the model to predict whether
BERT—are able to transfer what they’ve learned new, incoming transactions are fraudulent so
from a specific task to a more generalized set of you can flag them for closer study or block
tasks, including generating content. At this point, them outright.
you could ask a model to create a video of a car
going through a stop sign. — Convolutional neural network (CNN). CNNs
are a type of feed-forward neural network
Foundation models can create content, but they whose connectivity connection is inspired by
don’t know the difference between right and wrong, the organization of the brain’s visual cortex, the
or even what is and isn’t socially acceptable. When part of the brain that processes images. As
ChatGPT was first created, it required a great deal such, CNNs are well suited to perceptual tasks,
of human input to learn. OpenAI employed a large like being able to identify bird or plant species
number of human workers all over the world to help based on photographs. Business use cases
hone the technology, cleaning and labeling data include diagnosing diseases from medical
sets and reviewing and labeling toxic content, then scans or detecting a company logo in social
flagging it for removal. This human input is a large media to manage a brand’s reputation or to
part of what has made ChatGPT so revolutionary. identify potential joint marketing opportunities.

What is deep learning? 3

Here’s how they work: For more on deep learning, and neural networks
and their use cases, see our executive’s guide to AI.
• First, the CNN receives an image—for Learn more about McKinsey Digital.
example, of the letter “A”—that it
processes as a collection of pixels.
What is a foundation model?
• In the hidden layers, the CNN identifies Foundation models are deep learning models
unique features—for example, the trained on transformer network architecture: vast
individual lines that make up the letter “A.” quantities of unstructured, unlabeled data.
Foundation models can be used for a wide range of
• The CNN can then classify a different tasks, either out of the box or adapted to specific
image as the letter “A” if it finds that tasks through fine-tuning. Fine-tuning involves a
the new image has the same unique relatively short period of training on a labeled data
features previously identified as making up set, which is typically much smaller than the data
the letter. set on which the model was initially trained. This
additional training allows the model to learn and
— Recurrent neural network (RNN). RNNs are adapt to the nuances, terminology, and specific
artificial neural networks whose connections patterns found in the smaller data set. Examples of
include loops, meaning the model both moves foundation models include DALL-E 2, GPT-4, and
data forward and loops it backward to run again Stable Diffusion.
through previous layers. RNNs are helpful for
predicting a sentiment or an ending of a
sequence, like a large sample of text, speech, or What is a large language model?
images. They can do this because each Large language models are a class of foundation
individual input is fed into the model by itself as models that can process massive amounts of
well as in combination with the preceding input. unstructured text. These models can learn the
relationships between words or portions of words,
Continuing with the banking example, RNNs can also known as tokens. This enables large language
help detect fraudulent financial transactions models to generate natural language text, or
just as feed-forward neural networks can, but in perform tasks like summarization or knowledge
a more complex way. Whereas feed-forward extraction. Google’s Gemini runs on a large
neural networks can help predict whether one language model called LaMDA.
individual transaction is likely to be fraudulent,
recurrent neural networks can “learn” from the Learn more about McKinsey Digital.
financial behavior of an individual—such as a
sequence of transactions like a credit card
history—and measure each transaction against
the person’s record as a whole. It can do this in
addition to using the general learnings of the
feed-forward neural network model.

What is deep learning? 4

Which sectors can benefit from call service detects that a caller is getting upset,
machine learning and deep learning? the system can reroute to a human operator
or manager.
McKinsey collated more than 400 use cases of
machine and deep learning across 19 industries
Learn more about McKinsey Digital. And check out
and nine business functions. Based on our
deep learning–related job opportunities if you’re
analysis, we believe that nearly any industry can
interested in working with McKinsey.
benefit from machine and deep learning. Here
are a few examples of use cases that cut across
Articles referenced:
several sectors:

— Predictive maintenance. This use case is crucial — “Technology’s generational moment with
generative AI: A CIO and CTO guide,” July 11,
for any industry or business that relies on
2023, Aamer Baig, Sven Blumberg, Eva Li,
equipment. Rather than waiting until a piece of
Douglas Merrill, Adi Pradhan, Megha Sinha,
equipment breaks down, companies can use
Alexander Sukharevsky, and Stephen Xu
predictive maintenance to project when
maintenance will be needed, thereby reducing
potential downtime and lowering operating — “A new and faster machine learning flywheel
costs. Machine learning and deep learning have for enterprises,” March 10, 2023, Medha
the capacity to analyze large amounts of Bankhwal and Roger Roberts
multifaceted data, which can increase the
precision of predictive maintenance. For — “Deep learning in product design,” December
example, AI practitioners can layer in data from 14, 2022, Mickael Brossard, Jacomo Corbo,
new inputs, like audio and image data, which Marie Klaeyle, and Bill Wiseman
can add nuance to a neural network’s analysis.
— “An executive’s guide to AI,” November 17,
— Logistics optimization. Using AI to optimize 2020, Michael Chui, Brian McCarthy, and
logistics can reduce costs through real-time Vishnu Kamalnath
forecasts and behavioral coaching. For
example, AI can optimize routing of delivery — “Deep learning’s origins and pioneers,”
traffic, improving fuel efficiency and reducing May 8, 2018
delivery times.
— “Notes from the AI frontier: Applications and
— Customer service. AI techniques in call centers value of deep learning,” April 17, 2018, Michael
can help enable a more seamless experience Chui, James Manyika, Mehdi Miremadi,
for customers and more efficient processing. Nicolaus Henke, Rita Chung, Pieter Nel, and
The technology goes beyond understanding a Sankalp Malhotra
caller’s words: deep learning analysis of audio
can assess a customer’s tone. If the automated

Get to know and directly engage with McKinsey experts on deep learning.
Aamer Baig is a senior partner in McKinsey’s Chicago office, where Alex Singla is the global leader of
QuantumBlack, AI by McKinsey, and a senior partner; Sven Blumberg is a senior partner in the
Dusseldorf office; Michael Chui is a partner at the McKinsey Global Institute and is based in the Bay
Area office; Alex Sukharevsky is a senior partner in the London office; and Bill Wiseman is a senior
partner in the Seattle office.

Copyright © 2024 McKinsey & Company. All rights reserved.

What is deep learning? 5