Deep learning banner: box with smaller boxes emanating out the top

Deep learning is a subset of machine learning that deals with algorithms that mimic the function of the brain, called artificial neural networks, which learn from large sets of data. It is called “deep” learning since it uses multiple layers in a network, making it deeper than other more simple subsets of machine learning. 

Deep learning is a modern variation on early work which showed that a linear perceptron could not be a universal classifier, but that a network with a nonpolynomial activation function and one hidden layer could be. 

In order to make practical applications possible and optimize implementation, deep learning uses an unbounded number of layers of bounded size, while maintaining theoretical universality under normal conditions. Layers in deep learning networks are permitted to be heterogeneous in order to be efficient, trainable, and understandable.  

Why is deep learning important?

Deep learning may seem mysterious, but it is clearly powerful. For example, Google’s search engine, voice recognition systems, and self-driving cars are all heavily reliant on deep learning. Understanding how deep learning works in comparison with other types of machine learning makes it easy to see why it is so important. 

For example, traditional supervised machine learning systems classify images based on labels that the machine learning model has been taught. It will compare the image it sees to the labeled images it learned from. Deep learning differs from this, since labels are not required, meaning it functions as unsupervised learning. In unsupervised learning, the models will learn to classify images of the same type without ever being taught the labels that go with each image.

Your imagination can run wild with the possibilities for using this type of learning in the real world. Let’s dive into more about how deep learning works and then discuss its uses.

How deep learning works

In order to understand how deep learning works, you must first understand the meaning of and differences between a few terms. 

First, artificial intelligence (AI) refers to the replication of human intelligence within computers. For example, the beginning of AI research consisted of attempts to replicate human tasks, such as playing a game. Researchers programmed a number of rules that the computer would have to respect, so it made decisions based on those rules with a specific list of possible actions.

Next, machine learning (ML) is the ability of a computer to learn from large datasets rather than hard coded rules. ML means that computers learn by themselves, which is possible thanks to modern computers that can process huge amounts of data.

Supervised learning is a type of ML that involves labelled datasets with inputs and expected outputs. A supervised learning ML model is trained to get specific outputs from the inputs, and readjusts its calculations until it doesn’t get any outputs wrong. For example, a weather predicting model is trained using historical weather data, where it is given pressure, humidity, wind speed, etc. as inputs and generates the temperature as an output. 

Unsupervised learning, as we mentioned earlier, is machine learning with no specified structure. Models are given datasets to learn from, but do not have expected outputs for each input. The model learns completely on its own and can therefore teach us which insights we can gain from the data. For example, an unsupervised learning model could be given user behavior data and tell us which types of users are most likely to purchase which products.

Deep learning works with both supervised and unsupervised learning, but what really sets deep learning apart from other types of machine learning is its neural networks. Neural networks are the core of how deep learning works.

Neural networks

Neural network in connected dot matrix

AI systems have neurons, just like the brain. Think of these neurons as a bunch of circles that are interconnected. These groups of neurons, or circles, are then separated into three layers: an input layer, a hidden layer, and an output layer. 

The input layer is exactly what it sounds like. It receives the input data, and then passes it on to the hidden layer. 

The hidden layer then performs mathematical calculations with the input data. Deep learning neural networks have multiple hidden layers that each perform different computations, which is what makes deep learning so deep. A lot can be done with the input data as it travels through the hidden layers.

The output layer provides the output data. Neural network outputs are often predictions or other insights that researchers hope to glean from the dataset.

Deep learning uses

The uses for deep learning are endless, but here are a few examples of ways that deep learning is being used now.

List of deep learning use cases

  • Image recognition: This use of deep learning not only aims to recognize and identify people and objects in images, but also aims to understand the content and context of the image as well. This is a step further than less sophisticated ML models that label simple images with given labels, as we discussed earlier.
  • Voice recognition: Voice search and voice assistants use this type of deep learning. Voice recognition is one of the most widely used deep learning applications, since almost every cell phone is equipped with voice-to-text capabilities.
  • Text generation: These models are created to learn how to spell, form sentences, punctuate, and even follow certain styles of text that they have learned. Neural networks learn the relationships between words in the input text and can then generate grammatically correct sentences.
  • Predictions: Like the examples from earlier about weather and user behavior predictions, deep learning can be used to predict pretty much anything. This technology is being used to predict everything from earthquakes to financial market trends.

Applications of deep learning

The deep learning uses discussed previously can be used in many different applications. The most sophisticated applications of deep learning combine multiple uses of deep learning together to provide a holistic artificial intelligence solution. Here are a few examples of deep learning applications.

  • Automatic translation: Deep learning powers translate tools like Google’s, which can translate text, spoken words, and even images with text in them into many other languages. This uses voice recognition, image recognition, and text generation functions of deep learning.
  • Automatic colorization: Tools that colorize images automatically use deep learning models that have learned to recognize the colors of images based on how they look in black and white. 
  • Healthcare solutions: Detecting cancer and other anomalies is getting easier and more accurate thanks to deep learning and image recognition. Other healthcare solutions powered by deep learning include the many mobile apps and devices that monitor certain aspects of users’ health and can make predictions based on user health data. Deep learning innovations are reshaping healthcare.

 

  • Self-driving cars: Of course full self-driving cars, like the cars Google, Uber, and Tesla are all working on are powered by deep learning, but so are the driver assistance features, like automatic braking, that are already available in cars on the market.

Deep learning is powerful, but hopefully it seems less mysterious now. If you’d like to learn more about deep learning, the Algorithmia Blog is a great place to browse!

Continue learning

Types of machine learning: supervised and unsupervised

Statistics and machine learning: what’s the difference?

Prediction modeling: how it’s used and how it works