What Is Deep Learning?
Deep learning is a type of machine learning that uses artificial neural networks to enable digital systems to learn and make decisions based on unstructured, unlabelled data.
In general, machine learning trains AI systems to learn from acquired experiences with data, recognise patterns, make recommendations and adapt. With deep learning in particular, instead of just responding to sets of rules, digital systems build knowledge from examples and then use that knowledge to react, behave and perform like humans.
Why deep learning matters
Data scientists and developers use deep learning software to train computers to analyse big and complex data sets, complete complicated and nonlinear tasks and respond to text, voice or images, often faster and more accurately than humans. These capabilities have many practical applications and have made many modern innovations possible. For example, deep learning is what driverless cars use to process images and distinguish pedestrians from other objects on the road or what your smart home devices use to understand your voice commands.
Deep learning matters because as data volumes increase and computing capacity becomes more powerful and affordable, companies across retail, healthcare, transportation, manufacturing, technology and other sectors are investing in deep learning to drive innovation, unlock opportunities and stay relevant.
How deep learning works
Deep learning works by relying on neural network architectures in multiple layers, high-performance graphics processing units deployed in the cloud or on clusters and large volumes of labelled data to achieve very high levels of text, speech and image recognition accuracy. All that power can help your developers create digital systems with something like human intelligence and streamline time to value by accelerating model training from weeks to hours.
For example, a driverless car model might require thousands of video hours and millions of images to train. Without deep learning, this level of training could not be done at scale.
What is a deep learning framework?
To make complex machine learning models easier to implement, developers turn to deep learning frameworks like TensorFlow or PyTorch. These frameworks help streamline the process of collecting data which can then be used to train neural networks. In addition, accelerators like ONNX Runtime can be used with these frameworks to accelerate training and inferencing models.
Training deep learning models
There are different strategies and methods for training deep learning models. Let us take a closer look at a few of them.
With supervised learning, an algorithm is trained on datasets that are labelled. This means that when the algorithm makes a determination about a piece of information, it can use the labels included with the data to check if that determination is correct. With supervised learning, the data that models are trained on must be provided by humans, who label the data before using it to train the algorithm.
With unsupervised learning, algorithms are trained on data that does not contain labels or information that the algorithm can use to check its determinations against. Instead, the system sorts and classifies the data based on the patterns that it recognises on its own.
With reinforcement learning, a system solves tasks using trial and error to make series of decisions in sequence and achieve an intended outcome even in an environment that is not straightforward. With reinforcement learning, the algorithm doesn’t use datasets to make determinations, but rather information that it gathers from an environment.
Deep reinforcement learning
When deep learning and reinforcement learning techniques are combined, they create a type of machine learning called deep reinforcement learning. Deep reinforcement learning uses the same trial-and-error decision making and complex goal achievement as reinforcement learning, but also relies on deep learning capabilities to process and make sense of large amounts of unstructured data.
What is deep learning used for?
Deep learning is used within businesses in a variety of industries for a wide range of use cases. Here are some examples of how deep learning is commonly used:
Image, speech and emotion recognition
Deep learning software is used to increase image, speech and emotion recognition accuracy and to enable photo searches, personal digital assistants, driverless vehicles, public safety, digital security and other intelligent technologies.
Streaming services, e-commerce retailers, and other businesses use deep learning models to drive automated recommendations for products, movies, music or other services and to perfect customer experiences based on purchase histories, past behaviour and other data.
Savvy businesses use deep learning to power text- or voice-activated online chatbots for frequently asked questions, routine transactions and especially for customer support. They replace teams of service agents and queues of waiting customers with automated, contextually appropriate and useful responses.
Personal digital assistants
Voice-activated personal digital assistants use deep learning to understand speech, respond appropriately to queries and commands in natural language and even crack wise occasionally.
The unofficial representative for AI and deep learning, self-driving cars use deep learning algorithms to process multiple dynamic data feeds in split seconds, never have to ask for directions and react to the unexpected—faster than a human driver.
Many businesses use open-source machine learning software to bring deep learning solutions to their organisations.
What are neural networks?
An artificial neural network (ANN) is a digital architecture that mimics human cognitive processes to model complex patterns, develop predictions and react appropriately to external stimuli. Structured data is required for many types of machine learning, versus neural networks, which are capable of interpreting events in the world around them as data that can be processed.
Whenever you read a report, watch a movie, drive a car or smell a flower, billions of neurons in your brain process the information through tiny electric signals. Each neuron processes inputs and the results are output to the next neuron for subsequent processing, ultimately and instantly producing a business insight, a chuckle, a foot on the brake or a little joy. In machine learning, neural networks allow digital systems to interpret and react to situations in much the same way.
An ANN is like a brain full of digital neurons, and while most ANNs are rudimentary imitations of the real thing, they can still process large volumes of nonlinear data to solve complex problems that might otherwise require human intervention. For example, bank analysts can use an ANN to process loan applications and predict an applicant’s likelihood of default.
What you can do with neural networks
In machine learning, neural networks are used for learning and modelling complex, volatile inputs and outputs, inferring unseen relationships and making predictions without data distribution restrictions. Neural network models are the foundation for many deep learning applications, such as computer vision and natural language processing, which can help support fraud protection, facial recognition or autonomous vehicles.
Most businesses rely on forecasting to inform business decisions, sales strategies, financial policies and resource utilisation. But the limitations of traditional forecasting often make it difficult to predict complex, dynamic processes with multiple and often hidden underlying factors, such as stock market prices. Deep learning neural network models help expose complex nonlinear relationships and model unseen factors so that businesses can develop accurate forecasts for most business activities.
Common neural networks
There are dozens of different types of AI neural networks and each is suitable for different deep learning applications. Use an ANN that is appropriate for your business and technology requirements. Here are some examples of common AI neural networks:
Convolutional neural network (CNN)
Developers use a CNN to help AI systems convert images to digital matrices. Used primarily for image classification and object recognition, CNNs are appropriate for facial recognition, topic detection and sentiment analyses.
Deconvolutional neural network (DNN)
If complex or high-volume network signals get lost or convoluted with other signals, a DNN will help find them. DNNs are useful for processing high-resolution images and optical flow estimates.
Generative adversarial network (GAN)
Engineers use a GAN to train models how to generate new information or material that mimics the specific properties of the training data. GANs help the models distinguish subtle differences between originals and copies to make more authentic copies. GAN applications include high-fidelity image and video generation, advanced facial recognition and super resolution.
Recurrent neural network (RNN)
An RNN inputs data to hidden layers with specific time-delays. Network computing accounts for historical information in current states and higher inputs don’t change the model size. RNNs are a good choice for speech recognition, advanced forecasting, robotics and other complex deep learning workloads.
Transformers are designed to handle sequential input data. However, they are not restricted to processing that data in sequential order. Instead, transformers use attention—a technique that allows models to assign different levels of influence to different pieces of input data and to identify the context for individual pieces of data in an input sequence. This allows for an increased level of parallelisation, which can reduce model training times.
Machine learning vs. neural networks
Although neural networks are considered a subset of machine learning, there are some significant differences between neural networks and regular machine learning models.
For one, neural networks are generally more complex and capable of operating more independently than regular machine learning models. For example, a neural network is able to determine on its own whether its predictions and outcomes are accurate, while a machine learning model would require the input of a human engineer to make that distinction.
Additionally, neural networks are structured so that the neural network can continue to learn and make intelligent decisions all on its own. Machine learning models, on the other hand, are limited to decision-making based only on what it has specifically been trained on.
Get started with deep learning on Azure
Support image recognition with neural networks or optimise natural language processing with fast, easy, accurate BERT NLP models. Explore how to build, train and deploy all types of deep learning models with Azure Machine Learning.
Explore machine learning for data scientists
Learn how to build cloud-scale machine learning solutions on Azure and explore machine learning tools for data scientists and machine learning engineers.