Machine learning has changed the way we interact with computers in our daily lives. From shopping recommendations to robotic vacuum cleaners, machine learning algorithms and artificial intelligence allow computers to make smarter decisions from real-world data.
Deep learning represents the next step in artificial intelligence evolution. It is a type of machine learning that can use vast amounts of data to make even more intelligent decisions -- and even new discoveries. As AI expert Andrew Ng said, “[Deep learning] is our best shot at progress towards real artificial intelligence.”
How does deep learning work? How is it being used right now to improve our lives? And how can you learn more about deep learning? We’ll answer these questions in this article.
What’s “Deep” About Deep Learning?
Deep learning is a specific type of machine learning that focuses on the use of neural networks combined with vast amounts of data and large amounts of computing power to build even more intelligent machines.
To understand what makes deep learning so powerful, let’s look at some of the techniques that it uses, and how they make it different from other types of machine learning and AI.
Deep Learning Uses Multiple Neural Networks for Faster Learning
An artificial neural network is a system that aims to mimic the way human brains process information. Statistical algorithms are interconnected in a similar way to how the neurons in the human brain are connected, passing signals to each other in the form of numbers. This system can then be trained to perform tasks by feeding it inputs that result in a known output. It “learns” by adjusting itself until the output for a given input is correct.
For example, a neural network designed for image recognition could be trained to recognize a cat by feeding it images that are labeled “cat” and “not cat”. Eventually it will “learn” which pictures contain cats and be able to recognize unlabeled pictures correctly. This process is called “supervised learning”.
In deep learning, multiple neural networks are combined into layers that can communicate with each other. Each layer handles a particular part of the required task. In the image recognition example, one layer could handle recognizing edges, which feeds another that recognizes lines, then another that recognizes shapes, and so on. While the whole system may need training, each individual layer can use unsupervised learning -- that is, self-training -- to independently improve its ability to perform its task.
This is the “deep” part of deep learning -- the way the layers interconnect to build an even more intelligent system that improves over time.
Deep Learning Uses More Data to Learn
Another feature that sets deep learning apart from other types of machine learning is its use of vast amounts of data. While all machine learning requires data both for supervised learning and to improve its performance over time, deep learning systems may have access to hundreds of terabytes or even petabytes (1,024 terabytes) of data -- much more than a single computer typically stores.
In addition, deep learning systems use the data created each time the system processes an operation to improve itself. Because of the complexity of the system, it may generate an enormous amount of data in a short amount of time. This is why deep learning systems are usually built in large data centers that have access to large amounts of data storage.
Deep Learning Uses Powerful Computing Resources
Deep learning is different from other types of machine learning in the type and amount of computing resources that are used. A deep learning system is typically designed in a way that allows it to run on multiple computers at the same time. This is called clustered computing or distributed computing, and it allows systems to be built that have supercomputer levels of performance but use inexpensive, off-the-shelf CPUs and hardware.
In recent years, graphics processing units (GPUs) have also been used for deep learning systems. General-purpose computing on GPU (GPGPU) technology like NVIDIA’s CUDA architecture have made it possible for non-graphics programs to take advantage of the parallel processing capabilities of GPUs for machine learning. As with CPUs, GPUs running in multiple physical computers can be combined into clusters or distributed systems to build even more powerful deep learning systems.
What Deep Learning is Used For
The last few years have seen an explosion in deep learning and AI-based products. While computers used to only be able to guess at what we wanted, deep learning has changed the way we interact with our computers, phones, tablets and even TVs.
Deep Learning Powers Voice Recognition
If you’ve said “OK Google”, “Hey Siri”, or “Alexa” recently, you’ve used deep learning. Google, Apple and Amazon have invested heavily in using deep learning to build their voice assistant platforms. Voice recognition has been a challenge for computers for a long time, due to the many variations in dialects, accents and voice types. Most early voice recognition and dictation systems required extensive training, where the user had to read specific words and phrases so that the computer could “learn” their voice.
With deep learning, however, no training is required. These assistants can recognize a wide variety of voices and accents without needing to be trained by the user. Deep learning makes this possible by allowing the systems to continually improve themselves as millions of people use them every day around the world.
Voice assistants show another piece of the deep learning puzzle that’s important to understand. As mentioned, deep learning requires large amounts of data storage and compute resources -- far more than could be built into a phone or smart speaker. So these voice assistants connect via the internet to servers that actually handle the voice recognition and command processing. Your phone or smart speaker really just serves as a connection point to a much larger deep learning system.
Image Recognition and Processing with Deep Learning
Advanced image recognition and photo processing is another area that deep learning has impacted. As mentioned before, neural networks can be trained to recognize objects in images. Deep learning enables even more advanced image recognition. For example, a deep learning system can be trained not just to recognize that a photograph contains a dog, but the specific breed of the dog based on its appearance.
Services like Google Photos use deep learning to make huge photo catalogs searchable by text -- for example, you can search your photos for “food” or “blue cars” and it will return images matching your search.
Deep learning can also be used to enhance or repair photos. For example, deep learning can be used to blur backgrounds or add other effects to photos automatically. Or it can be used to make inferences about the contents of faded or damaged photographs, allowing us to see modern-looking depictions of historical scenes.
Deep Learning Speeds Pharmaceutical and Medical Research
A more significant application of deep learning is in the research and discovery of new medications and treatments. Deep learning can be used to predict how new drug formulations will impact both targeted disease cells (for example, a specific kind of virus) and other cells in the body, which may cause side effects. This could speed up the rate at which new medicines are discovered and give us better information about how effective they will be.
Deep learning has also been used in genetic research to better understand which genes do what. It can also be used in bioinformatics -- the study of data to make better health decisions. It can be used, for example, to analyze data from smart wearables to make better health decisions, or to scan medical images for signs of cancer or other problems.
These are just a few of the areas that deep learning has or will have an influence on. If it really is “our best shot at… real AI”, we will continue to see it being used more and more in business, technology, medicine, and our everyday lives.
Learning More About Deep Learning
At first, deep learning can seem like an intimidating subject to learn. It’s true that it is a very broad and complex field, and that some very intelligent people are working on it every day. But this doesn’t mean that you can’t get started learning about it and even working toward a career in it.
As mentioned, deep learning is a specific application of machine learning. In SoloLearn’s free Machine Learning course, you’ll use Python to explore the basics of machine learning, including classification, decision trees, and even neural networks. These concepts will put you well on your way to understanding deep learning.
Because deep learning relies heavily on data, you may also want to explore the Data Science with Python course. These data analytics concepts are critical to understanding how data is used in machine learning and AI.