What is deep learning?

Technology continues to advance at incredible rates. From self-driving vehicles to supercomputers and autonomous drones, the kind of world depicted in I, Robot, is becoming our reality – for better or worse. What’s driving this technology? Artificial intelligence (AI) and its many subcategories, including machine and deep learning.

While AI is an enormous field that encompasses all machines that can execute tasks that otherwise require human intellect, and machine learning is the process by which machines can learn new tasks autonomously, deep learning is a different beast entirely. Deep learning is a subset of both AI and machine learning and attempts to build the first machines capable of human-like thought. It relies on algorithms to process information and ‘learn’ new skills in a way that mimics how the human brain learns.

How deep learning works

So, if deep learning is about training machines to process code like the human brain, how does it work?

Deep learning involves the development of complex, layered algorithms, called artificial neural networks, that allow machines to continually learn from large libraries of data. In order for an algorithm to be considered deep, it usually requires at least six layers, but often more. Each layer of code is responsible for finding a specific pattern in the dataset. For example, a deep learning algorithm that seeks to suggest new music for a company like Spotify would have several layers. One layer might look specifically at previous genres played. Another layer might look at song likes, dislikes, and yet another at songs that were skipped or repeated.

As the dataset grows, the algorithms become even more powerful, as the algorithms are built to ‘learn’ independently as new data becomes available. This feature is unique to deep learning and makes it scalable and very much in line with how the human brain adapts to new information.

Types of deep learning layers

If deep learning sounds complicated, that’s because it is. Deep learning algorithms rely on artificial neural networks for computation, and each layer serves a different role. Layers like the Multilayer Perceptron Neural Network (MLPNN) is a learning algorithm that searches for dependencies and correlations between input and output data. The backpropagation layer is the primary training layer and calculates the weighted gradient descents. It also updates backward from output toward input and is key to the training of the neural network.

Other layers include the Convolutional Neural Network (CNN), which mostly analyzes visual data; the Recurrent Neural Network (RNN), which works mostly to identify patterns and trends and predict upcoming scenarios; Long Short-Term Memory (LSTM) to stabilize weighted gradients; Generative Adversarial Networks (GAN) for new data generation; the Restricted Boltzmann Machine (RBM) Model for binary factor analysis; and Deep Belief Networks (DBN) for generating new generative models based on probability.

The more complex the deep learning algorithm, the more layers it will have, and also, the more effective it is.

Deep learning is the future of AI

Deep learning is still an early development, but it’s certainly one of the more powerful techniques in use today to make autonomous, self-thinking machines. While today it’s used in applications such as virtual assistants, marketing algorithms, facial recognition software, and chatbots, this same technology will also power self-driving cars, autonomous data-gathering drones, and so much more in the not-so-distant future.

If you’re interested in learning more about deep learning, check out the book Deep Learning (MIT Press), a foundational text in the field. You can also keep up with the Google Brain Team and its ongoing deep learning research projects.









What is artificial intelligence (AI)?

What is machine learning?