Artificial Intelligence: Deep Learning vs. Machine Learning Explained
Artificial Intelligence: Deep Learning vs. Machine Learning Explained
As artificial intelligence, or AI, becomes more commonplace in our everyday lives, it is important to understand just exactly what the buzzwords artificial intelligence, deep learning, and machine learning are all about, and what makes them different from each other. In the following paragraphs we will do our best to explain what they are and how they are different in simple, layman’s terms.
You can think of these three distinct, yet related terms like nesting dolls.
The outermost doll is artificial intelligence or AI. AI is the study of ways to create programs that can solve problems in a creative manner, rather than a pre-programmed solution.
Inside the AI doll is machine learning. Machine learning is a subset of artificial intelligence, and describes a system that automatically learns and improves as it gains experience. Machine learning uses different algorithms to facilitate learning and does not need to be specially programmed to complete a specific task. Machine learning allows for the recognition of patterns in data and the ability to make predictions.
The innermost doll is deep learning, which is sometimes referred to as deep neural learning or a deep neural network. Deep learning is a subset of machine learning that describes the use of multi-layered neural networks that are modeled after a human’s brain to learn from, analyze and solve problems from unstructured and unlabeled data. These networks are made up of multiple different, yet related processes that communicate with each other to ultimately solve a problem.
How does a machine learn?
Perhaps not surprisingly, computers learn in much the same way that humans do. Let’s examine how machine learning applies to recognizing a basketball in an image. As a human, we gather upwards of 90% of our information visually, so we’ll start there.
As humans, we learn what a basketball is by being told that the round, orange ball with black ribs we are holding is a basketball. Having been shown one for the first time, we would be somewhat accurate at picking out other basketballs from a pile of objects, but we would not be perfect. We’d need to see and experience more basketballs so we could accurately differentiate them from volleyballs, tennis balls or even frisbees. After all, a volleyball is a large sphere that bounces, a tennis ball has ribs, and a frisbee is round when viewed from the top. Without more examples of a basketball, we can’t be 100% sure that the ball we are looking at is indeed a basketball.
Machine learning works in the same way; we label an image of a basketball as such, the computer examines it, guesses at what attributes combine to make the object a basketball and remembers them. We then feed it thousands, if not hundreds of thousands of other images that are labeled as basketballs, and the computer refines its understanding of what makes up a basketball until it can accurately recognize one when asked to do so.
What makes learning “deep”?
Deep learning differs from machine learning in how the information is processed. Both machine learning and deep learning require an input and produce an output by utilizing algorithms. The difference lies in how they move through the process. Machine learning uses an orderly or linear approach, meaning it follows a specific path to the answer every time, whereas deep learning is non-linear.
Machine Learning Explained
Let’s jump back to our machine learning basketball recognition example. Here, we gave the machine an input of a photo with a basketball somewhere in it (step 1). The machine determines the important features (step 2) of the basketball (shape, color, presence of ribs, etc.), then classifies the image as containing a basketball or not (step 3), then tells us the result (step 4).
Deep Learning Explained
With deep learning, the process is simplified to three steps; (1) input, (2) feature extraction and classification, and (3) output. The deep learning approach combines the feature extraction and classification step by using multiple neural layers to complete these tasks. Each of these layers pass weighted information to the next layer, where it is acted upon. For example, a neuron may be looking for the presence of the color orange while another looks for ribs, etc. This repeats until the output layer is reached, and the results of the process are reported.
What makes something artificially intelligent?
The concept of artificial intelligence dates back to the 1950s when a group of researchers convened to work out how to make a model of the human brain on a computer. While they were unable to meet that goal, they did define three key abilities a machine should have to be considered artificially intelligent: learning, natural language processing, and creativity.
Key AI Ability: Learning
In order to be considered AI, a machine must be able to learn. What this means is that AI must be able to acquire greater knowledge on a subject by studying, experiencing or being taught about it. Take your email spam detector as an example of learning in AI.
When you mark a message as spam, your filter learns from this action so that future spam messages do not make it into your inbox. Perhaps it looks at the sender and remembers that they spend spam, or perhaps it looks at the links in the email and decides that links formatted in a similar manner are probably spam. It even looks at the sentence structure and grammar of the email to determine how a spammer writes. Your marking of items as spam or not spam serves to reinforce this machine learning.
Key AI Ability: Natural Language Processing
Before the advance of AI, communicating with a computer was a very regimented process. Entire languages have been developed that allow us to tell a computer what to do and the outputs of computers were limited to what they were programmed to say.
In order to be artificially intelligent, a computer should be able to converse in a human manner — to the point that a human would have difficulty determining if they are conversing with a real human or computer. The Turing Test is a popular way to determine if a machine has achieved artificial intelligence. During a Turing Test, a human asks questions and both a human and computer answers them. After questioning, the human is asked to decide which respondent was human and which was a computer. If the human is unable to answer, then the computer is considered to be artificially intelligent.
Key AI Ability: Creativity
Computers are designed to follow and execute instructions in a predictable manner, and nothing more. Any deviation from programming is considered an error. Artificially intelligent programs are not limited to specifically programmed processes, and can make use of other input or data to optimize a process. To better explain this, let’s use the maps app on your phone as an example.
Back in the 1990s we all used MapQuest to pre-plan routes, printing out paper instructions to follow before a trip. These instructions were typically based on the shortest route between points A and B, and if you deviated from that route, there was no way for your paper instructions to adjust. Today, our maps apps creatively re-route us based on a wide range of inputs. Are other maps users going especially slow ahead? Was an accident reported on our path? Is there construction on this road? Perhaps it’s best to take a short detour.
In Conclusion
To put it most simply, artificial intelligence, machine learning and deep learning are all directly related. Machine learning is a subset of artificial intelligence, and deep learning is a subset of machine learning. Machine learning differs from deep learning by how the program is processed. Machine learning is given an input, looks for specific features in the image, determines if the image contains those features and lets you know if the features were found. Deep learning, on the other hand, takes in input and processes/classifies features at the same time with a neural network similar to the human brain.
