WHAT IS
What is deep learning?
Deep learning is a subset of artificial intelligence (AI) that mimics a brain’s neural
networks to learn from large amounts of data, enabling machines to solve complex
problems.
Published on June 11, 2024
AI
Deep learning definition
Deep learning definition
Deep learning is a type of machine learning that enables computers to process
information in ways similar to the human brain. It's called "deep" because it involves
multiple layers of neural networks that help the system understand and interpret data.
This technique allows computers to recognize patterns and manage complex tasks, such
as translating languages and driving cars autonomously. Similar to how humans learn
from experience, these systems improve their skills and accuracy over time by analyzing
vast amounts of data, without needing manual updates from humans.
Understanding neural networks
From theory to Perceptron
In the 1940s, Warren McCulloch, a neuroscientist, and Walter Pitts, a mathematician,
collaborated to create the first artificial neural network concept. Their goal was to
understand how the brain could produce complex thought patterns from the simple
binary responses of neurons. They introduced a model of the neuron, which they
believed could mimic the brain's ability to perform complex calculations using binary
logic.
In the neural network model developed by McCulloch and Pitts, inputs act like the
electrical impulses a neuron receives. If some inputs are more crucial for a specific result,
the model emphasizes these through greater weight. When these weighted inputs
exceed a certain level, the neuron activates; if not, it remains off. This basic on-off
mechanism enabled their model to mimic simple brain-like decision-making processes,
setting the stage for deep learning's evolution.
In 1957, the introduction of the Mark I Perceptron, a room-sized machine built by
computer scientist and psychologist Frank Rosenblatt, showcased the first practical use
of artificial neurons. This device used photocells and artificial neural networks to
recognize and categorize images, demonstrating the effectiveness of McCulloch and
Pitts' ideas. Rosenblatt's Perceptron not only confirmed that machine learning could
work but also paved the way for the development of today's more sophisticated deep
learning technologies.
How does deep learning work?
Deep learning works by using the process of prediction to determine which algorithms in
their neural networks are the most successful at producing outputs that meet human
expectations. Then, the networks use backpropagation to refine those algorithms so that
their rate of success improves. Here’s an example:
Imagine you're teaching a computer to recognize different genres of music. The neural
network analyzes thousands of music files, gradually learning to notice features like
instrumentation, beats, and chord progressions. When it makes a prediction, like
identifying a piece as a rock song, and is then told whether it's correct, it uses a method
called backpropagation to adjust its algorithm.
This is like learning from mistakes. For example, if the computer mistakes a classical
piano sonata for a rock song, it learns from this error, refining its ability to distinguish
between classical and rock songs in future predictions. Over time, this process enables
the artificial neural network to make highly accurate predictions, turning it into a powerful
tool for everything from recommending movies based on what you like to enabling self-
driving cars to interpret road signs and signals.
A deeper dive into deep neural network layers
This list explains the essential components of a deep neural network and the general
order in which they function. However, neurons, activation functions, and regularization
techniques are not isolated steps, but rather features that operate throughout the
network and its learning process.
1. Input layer
The input layer is the gateway into the network, where each neuron represents a unique
feature of the input data. This layer's primary function is to receive the raw data and pass
it to the subsequent layers for further processing.
1. Neurons (nodes)
Neurons, or nodes, are the fundamental processing units of a neural network. Each
neuron receives input, processes it (using a weighted sum and then applying an
activation function), and sends the output to the next layer.
1. Activation functions
These are like the decision-makers in a neural network, helping it determine what to
learn and what to ignore. They add a kind of flexibility to the network, allowing it to
capture and learn complex patterns. Common activation functions include sigmoid,
ReLU (rectified linear unit), and tanh.
1. Weights and biases
Weights are parameters within the network determine the influence of input data on the
outputs within the network's layers. Along with weights, biases ensure that activation
functions can produce non-zero outputs, enhancing the network's ability to activate and
learn effectively.
1. Hidden layers
Situated between the input layers and output layers, hidden layers perform the bulk of
computations within a neural network. They’re called "hidden" because unlike the input
and output, they don’t interact with the external environment. The complexity and
capability of a neural network are largely determined by the number and architecture of
hidden layers.
1. Output layer
This is the final layer in a neural network. It presents the results, transforming the
information from the hidden layers into a format that solves the task at hand, such as
classification, regression, or any other type of prediction.
1. Loss function
The loss function, or cost function, quantifies the difference between the predicted
outputs and actual outputs. Minimizing this function is the goal of training, enabling the
model to predict more accurately.
1. Optimization algorithms
These algorithms fine-tune the model to improve its accuracy over time. They tweak the
weights and biases to reduce errors during predictions. Some popular methods include
stochastic gradient descent, Adam, and RMSprop.
1. Backpropagation
This deep learning algorithm is crucial because it helps the model learn and improve
from its mistakes. It figures out how changes to the model's weights affect its accuracy.
Then, it adjusts these settings by tracing errors backward through the model to make it
better at making predictions.
1. Regularization techniques
Models often learn the training data too closely, causing them to not perform as well on
new data (known as overfitting). To adjust for this, techniques like L1 and L2
regularization and batch normalization are used to fine-tune the size of weights and
speed up the training process.
1. Batch normalization
This technique normalizes the inputs of each layer, aiming to improve the stability,
performance, and speed of the neural network. It also helps in reducing the sensitivity to
the initial starting weights.
1. Dropout
Another regularization method, dropout randomly ignores a set of neurons during
training. This helps to reduce overfitting by preventing the network from becoming too
dependent on any single neuron.
Common applications of deep learning
Deep machine learning has come a long way since the Perceptron. Instead of needing to
install room-sized machines, organizations can now create deep learning solutions on
the cloud. The ability of today’s deep neural networks to handle complex datasets make
them valuable tools across diverse sectors, opening new avenues for innovation that
were once considered futuristic.
Automotive
Deep learning allows vehicles to interpret sensor data for navigation. It also improves
driver assistance systems, with features like hazard detection and collision avoidance,
and contributes to better vehicle design and manufacturing.
Business operations
Conversational AI chatbots and virtual assistant copilots are popular business deep
learning applications. They reduce human error by automating manual tasks, accelerate
data analysis and decision-making, and make it easier to find information stored across
different systems.
Finance
Algorithmic trading powered by deep learning is used to analyze market data for
predictive insights and identifies complex patterns to enhances fraud detection. Deep
learning also aids in risk management, evaluating credit risks and market conditions for
more informed decision-making.
Discover more about AI technology in finance
Healthcare
Deep learning algorithms can help improve diagnostic accuracy and detect anomalies
like tumors at early stages from medical images. There are also opportunities for drug
discovery by predicting molecular behavior, facilitating the development of new
treatments.
Manufacturing
Predictive maintenance uses the Internet of Things and deep learning to anticipate
machinery failures, minimizing downtime. Visual inspection systems trained on extensive
image datasets can enhance quality control by identifying defects.
Discover more about AI technology in manufacturing
Media and entertainment
The entertainment industry uses deep learning applications to power content
recommendations on streaming platforms, and to help creators develop realistic CGI
and compose music using generative AI. It also analyzes viewer preferences, helping
creators tailor content and predict future trends.
Retail
Deep learning has revolutionized retail customer experiences with personalized product
recommendations. It also improves inventory management by using predictive analytics
to forecast demand and optimize stock levels.
Discover more about AI technology in retail
Supply chain
Logistics operations are using deep machine learning to optimize delivery scheduling by
identifying traffic disruptions in real time. Deep learning also enhances demand and
supply forecasting accuracy, enabling proactive strategy adjustments.
Deep learning benefits and challenges
While the benefits of deep learning are truly impressive, the complexity of this
technology brings challenges, too. And because deep learning solutions require
considerable planning and resources, it’s critical that organizations establish clearly
defined goals and responsible AI practices prior to designing and deploying this
technology.
Benefits Challenges
High accuracy in tasks like
Requires large datasets for training
image and speech recognition
Computationally intensive, needing
Ability to process and analyze
significant processing power
vast amounts of data
Can be a “black box,” making it
Improves over time as it's
difficult to understand models’
exposed to more data
decision processes
Automates feature extraction,
Susceptible to perpetuating unfair
reducing the need for manual
biases when training data is faulty
intervention
Needs continuous data and
Enables personalized
monitoring to maintain performance
experiences in services and
over time
products
Deep learning vs. machine learning vs. AI
There are some key differences between traditional machine learning and deep learning:
Machine learning relies on humans to manually identify and select the features or
characteristics of the data that are important for a task, such as edges in images or
specific words in text. This process of training requires a lot of expertise and effort.
Deep learning enables machines to automatically determine which features of the data
are most important for performing specific tasks. This is done by processing the raw data,
such as pixels in an image, through multiple layers of a neural network. Each layer
transforms the data into a more abstract form, building on the previous layer's output. As
the model is exposed to more data, it continuously refines these transformations to
improve accuracy and performance, becoming more effective over time.
AI vs. deep learning vs. machine learning example
If you’re unsure of the differences between AI, machine learning, and deep learning,
you’re not alone. Here’s a real-world AI vs. deep learning vs. machine learning example
about self-driving vehicles:
AI is the overarching technology used to give self-driving vehicles human-like intelligence
and autonomy. It includes machine learning and deep learning.
Machine learning is the subtype of AI that allows self-driving systems to learn and
improve from data without being specifically programmed for every scenario.
Deep learning is the specialized subtype of machine learning that processes and
interprets the complex inputs, including visual data from cameras, making sense of the
environment in real-time.
Deep learning vs. deep learning models
It’s also not uncommon to see the terms “deep learning” and “deep learning models”
used interchangeably, but there’s a nuanced difference between them:
Deep learning refers to the entire field of study. It encompasses the theories, techniques,
algorithms, and processes used to train artificial neural networks.
Deep learning models refer to the specific neural networks that have been designed and
trained to solve a particular problem or perform a specific task. Each model is unique,
tailored to its specific data, training, and task. A model's performance depends upon:
How well it’s been trained, including the quality and quantity of the data, and its
learning rate.
The design and computational power of the computer infrastructure it runs on.
What are deep neural networks?
Deep learning networks, often called deep neural networks, learn complex patterns in
large datasets by adjusting neural connections through training. There are several major
types: artificial neural networks, convolutional neural networks, recurrent neural
networks, generative neural networks, and autoencoders.
Deep neural network types
Artificial Convolutional Recurrent Gene
Feature/Type Neural Neural Neural Neura
Network Network Network Netw
General Image and Natural
Image
purpose, video language
gener
Primary use ranging from recognition, processing,
transfe
regression to image speech
augme
classification. classification. recognition.
Ability to
process
Use of
information in
convolutional
sequences, Ability
layers to
Key Simplicity and preserving gener
adaptively learn
characteristics versatility. information data s
spatial
from one step the in
hierarchies of
of the
features.
sequence to
the next.
Consis
A network of
Specialized for Designed for netwo
neurons/nodes
processing grid- sequential or (gene
Basic concept that simulate
like topology time-series discrim
the human
data. data. comp
brain.
game.
Capable of
learning long-
Powe
Flexible. Can High efficiency term
gener
be applied to a and dependencies
data i
Advantages broad performance in with
enhan
spectrum of tasks related to modifications
realism
tasks. visual data. like long
divers
short-term
memory.
Artificial Convolutional Recurrent Gene
Feature/Type Neural Neural Neural Neura
Network Network Network Netw
Difficulty in
May struggle
Requires a training over
with complex Trainin
significant long
pattern and m
amount of sequences
Challenges recognition in collap
training data for due to
raw, high- challe
optimal vanishing
dimensional manag
performance. gradient
data.
problem.
Convolutional
Chains of Two n
layers, pooling
Layers of fully repeating gener
Architectural layers, followed
connected units that create
features by fully
neurons. process a disc
connected
sequences. to eva
layers.
Excels at
Learn
handling
Handles a Efficiently gener
sequential or
Data handling wide range of handles spatial that is
time-
data types. data. indisti
dependent
from r
data.
Deep learning infrastructure requirements
Deep learning requires specialized computing and networking infrastructure to process
its complex models and massive datasets. It’s not practical to run deep learning models
on general computer hardware or networks, so many organizations adopt enterprise AI
platforms to meet the necessary requirements. Here are the main infrastructure
considerations:
High-performance GPUs
The backbone of deep learning infrastructure is high-performance graphics processing
units (GPUs). Originally designed for rendering graphics in video games, GPUs have
processing capabilities that make them well-suited for deep learning. Their ability to
perform multiple calculations simultaneously greatly reduces training time for models,
making them indispensable for modern AI research and applications.
Scalable storage solutions
The more data a model can learn from, the better its performance. This creates a need
for scalable and fast storage solutions that can handle petabytes of data without
creating bottlenecks in data retrieval. Solid state drives and distributed file systems are
commonly used to meet these demands, offering high-speed data access that keeps
pace with the computational speed of GPUs.
Efficient data processing frameworks
Frameworks and libraries such as TensorFlow, PyTorch, and Keras simplify the
development of deep learning models by providing pre-built functions, reducing the
need for coding from scratch. These tools not only speed up the development process
but also optimize the computational efficiency of training and inference, allowing for the
effective utilization of underlying hardware.
Cloud computing platforms
Cloud computing platforms play a pivotal role in making deep learning widely
accessible. They provide access to high-performance computing resources on-demand,
eliminating the need for significant upfront investment in physical hardware. These
platforms offer various services, including GPU instances, scalable storage, and machine
learning frameworks, making it easier for individuals and organizations to create and
deploy deep learning models.
Network infrastructure
Deep learning models are often trained across multiple GPUs and even across different
geographical locations, so a robust network infrastructure is crucial. High-bandwidth
connectivity ensures that data and model parameters can be efficiently transferred
between nodes in a distributed training setup, minimizing delays and optimizing the
training process.
Explore AI built for business
See how to enhance and better connect your people, data, and processes.
Learn more
FAQ
What is deep learning in simple words?
Deep learning, sometimes also called deep machine learning, is a type of artificial
intelligence that teaches computers to learn by example, much like humans do. It
uses a layered structure of algorithms called neural networks to process data,
recognize patterns, and make decisions.
What is an example of deep learning?
A notable example of deep learning is in medical imaging, where algorithms analyze
images like X-rays, MRIs, or CT scans to detect diseases such as cancer. By training
on vast datasets of medical images, these deep learning systems can identify subtle
patterns that might be missed by human eyes, assisting doctors in early diagnosis
and personalized treatment planning.
What are three types of deep learning?
1. Convolutional neural networks: A familiar example is the face unlock feature
on smartphones. Convolutional neural networks analyze the facial features from
the camera input to verify the user’s identity, allowing secure and quick access to
the device. This process involves the network learning from various images to
accurately recognize and confirm the user’s face.
2. Recurrent neural networks: Ideal for tasks involving sequences, such as
predicting the next word in a sentence. This makes them great for applications like
predictive text on your smartphone, where the network learns from the sequence
of your typing to suggest the next word you might type.
3. Autoencoders: A practical example is image compression, where autoencoders
reduce the size of images for storage or transmission and then restore them to
their original quality when needed. This process helps in reducing the space
needed to store images while maintaining their quality.
What is the difference between machine learning and deep
learning?
Machine learning refers to the broader concept of computers learning from data to
make decisions or predictions. Deep learning is a subset of machine learning that
uses neural networks with many, or “deep,” layers. The main difference is the depth
of learning; deep learning automatically discovers the most relevant data to be used
for learning, but machine learning requires the data to be specified manually.
Additionally, deep learning performs better with larger datasets, while traditional
machine learning can be more effective with smaller datasets.
Published on June 11, 2024
Share this article
Explore related content
AI AI implementation
Explore AI built for business
See how to enhance and better connect your people, data, and processes.
Learn more
Read more
BLOG BUSINESS CASE
People-machine blend: Companies warm up to
How CHROs can lead… solar as business value…
SAP Insights research center
Published on September 3, 2025 Published on September 3, 2025
WHAT IS
What is a self-
service portal?
Published on September 2, 2025
Quick links
Trending
About SAP
Site information
