Artificial intelligence (AI) is a vast, complicated field. We know, we work with AI every day, and while our work is very rewarding, it’s also extremely challenging. AI is also a diverse field and there are many different subsets, including deep learning and machine learning.
Sometimes these two important and closely related subfields are mixed up. And while they refer to related concepts there are some very important differences. In practice, deep learning and machine learning are used to solve different problems and challenges and each has its own unique strength and limitations.
In this article, we’ll outline what deep learning and machine learning are, why they’re different, and when you should use each. We’ll also examine their intertwined history so you can develop a more complete understanding.
Machine Learning and Deep Learning Defined
Let’s start by offering a text-book style definition for both:
Machine learning algorithms are created and implemented that allow for self-modification. In other words, no human input is needed. Instead, machine learning programs can analyze structured data and modify themselves based on said data.
Deep Learning uses artificial neural networks (ANN) to create multiple layers of algorithms that each offer interpretations of the data feed. Like machine learning, the algorithms can adjust themselves. The ultimate goal of deep learning is to imitate the human brain.
Based on the above definitions, you can see that machine learning and deep learning are closely related. In fact, deep learning uses machine learning. As such, some consider it a subset of machine learning.
However, deep learning represents such a major breakthrough that this classification may not do it proper justice. One could argue that machine learning is instead a component of deep learning.
Don’t worry if you don’t have a full grasp of machine learning and deep learning just yet. Let’s go over an example so you can understand how machine learning and deep learning work and when they are applied. Then we’ll dive into their history.
Machine Learning Versus Deep Learning Illustrated
Let’s assume you have a collection of photos of men and women, and you want to sort these photos by sex. With machine learning, you’d use structured data to sort the images. You can create structured data for the images based on specific features, such as long hair, facial structures, the width between eyes, and the like.
The machine learning algorithms will then use this structured data to do the sorting. If all goes well, your images will be sorted in no time. However, machine learning requires structured data to work. Often, most data is unstructured, and generally developers have to structure the data themselves.
Further, if the machine learning algorithm fails to properly sort the images, a human is going to have to step in and adjust the algorithms until they function properly. In practice, this means AI developers may spend a lot of time tweaking machine learning algorithms.
How Deep Learning Approaches the Same Problem
Now let’s say you use deep learning instead. With deep learning, you don’t need to provide structured data. Instead, the images can be fed through multiple layers of algorithms, which can identify specific features and then sort them. In this case, queries will be sent through various hierarchies, finding the appropriate identifiers to separate men and women.
With deep learning, there is often no need for humans to manually tweak the algorithms. Instead, the deep learning algorithms will adjust themselves until the appropriate outputs are matched. However, deep learning requires a vast amount of high-quality, sortable data.
So Which Is Better? It Depends on the Application
Based on the simplified account above, you might conclude that deep learning is the better technology. However, this isn’t necessarily true. True, deep learning is more complex, and in many, ways more powerful. Still, deep learning requires far more data than machine learning.
If the data is not available, the algorithms may not be able to function. In practice, deep learning is only applicable for complex queries and massive data sources. Machine learning, on the other hand, can work with more limited data sources and simpler queries.
The development of both machine learning and deep learning represented large leaps in the advance of artificial intelligence. In order to further our understanding of machine and deep learning, let’s put them both in a historical perspective.
The Evolution of Machine Learning and Deep Learning
Machine learning was actually one of the first breakthroughs in the field of artificial intelligence, which was a cutting edge field by the 1950s. Scientists wanted computers to emulate human thinking processes so that they could solve real-world problems.
The term “machine learning” was first coined in 1959 by one of the pioneers of AI, Arthur Samuel. He noted that machine learning was the “field of study that gives computers the ability to learn without being explicitly programmed.”
This would greatly reduce the need for human input. Samuel set out to build a computer program that could play checkers and eventually beat skilled humans. Most importantly, Samuel wanted the program to learn as it played.
Samuel’s efforts laid the foundation for the field of machine learning. Among other things, he outlined a method of rote learning for machines and a scoring function that allowed the machine to calculate the greatest probability of winning with any given checkers board.
AI Researchers Target More Complex Games
By the 1960s, machine learning algorithms were becoming quite competent, regularly beating humans. Some researchers set their sights on a higher goal: developing a machine learning program that could beat the world’s best chess players.
Chess is far more complex than checkers as there are many more outcomes. The first computer to beat a top chess player was IBM’s famous “Deep Blue.” However, Deep Blue didn’t use machine learning, but instead algorithms written by humans based on chess best-practices.
It’d take a major breakthrough for machine learning-enabled programs to beat world-class competitors in chess and other complex games. As you might have guessed, that breakthrough was deep learning. When it came to more complex games, machine learning simply lacked the “brain power” needed.
Sure, these programs could handle checkers, but chess and other complex games were a whole different story. With chess already conquered by Deep Blue, AI researchers at Google set their sights on an even more audacious goal: beating the world’s best “Go” Player.
For Deep Learning, It Was Time to Pass “Go”
Go is perhaps the most complicated popular board game in the world. This ancient game was invented in ancient China and has captivated societies ever since. Go’s rules are simple, you move black and white stones around the board, trying to conquer territory.
However, as far as AI is concerned Go, is more complex than checkers or even chess. Why? Consider this: there are more possible variations of stones on the game board than there are atoms in the universe. Such a vast amount of data is difficult for AI programs to deal with.
For AI researchers, creating an AI program that could teach itself to beat championship level Go players became the perfect benchmark. It didn’t come quickly. Machine learning and deep learning programs were trounced by top Go players for years.
In 2017, however, Google’s AlphaGo made a huge breakthrough, first beating South Korean Go Grandmaster Lee Sedol, then taking down Ke Jie, who at the time was considered the world’s best human Go player. This was a watershed moment for deep learning, so let’s look at how AlphaGo won.
How AlphaGo Learned and Then Won
So how was AlphaGo able to outmatch Ke Jie? The answer lies in Deep Neural Networks (DNN), which are a type of Artificial Neural Network (ANN) that allows for multiple input and output layers. The neural networks are able to take a description of the Go board as an input. Then, the board is processed through multiple network layers that contain millions of artificial “neurons.”
One network is established as the “policy network”, and it selects the next move to play. The “value network”, on the other hand, predicts the winner of the game. The multiple layers and dual networks allowed AlphaGo to do a far better job analyzing and predicting the fast-evolving Go games than previous AI Go programs.
With AlphaGo, the search tree so common in many AI applications was not needed. Indeed, due to the sheer number of outcomes, using a search data tree to build a champion-caliber Go program would be next to impossible.
AlphaGo did use machine learning as part of its learning process, however. It taught itself and adjusted itself over time. However, the deep neural networks were the key to finally building a computer program capable of beating top Go players.
Both Machine Learning and Deep Learning Will Be Used in the Future
Deep learning marks a dramatic advance for artificial intelligence and machine learning.
However, both of these technologies will be used frequently in the years ahead. When it comes to solving certain problems, especially with limited resources, machine learning may be a better choice.
Still, deep learning points to the future of artificial intelligence. As the field evolves, human intervention will likely become less necessary and AI will be able to tackle increasingly difficult tasks. As AI experts ourselves, we’ll be keeping a close eye on developments in both subfields.