Machine learning and deep learning are two subsets of artificial intelligence (AI) that are sometimes misunderstood. These terms may be used interchangeably and can lead to a bit of confusion. So how does deep learning differ from machine learning?
The simplest answer is that deep learning is just a method for implementing machine learning. Deep learning, machine learning, and AI are all part of the same family. AI is the grandparent, machine learning is the parent, and deep learning is the child. Each subset grew from the previous one, inheriting its traits and introducing new and improved capabilities.
A 60 Year History
The first known reference to the phrase “Artificial Intelligence” was in a proposal submitted in 1955 by John McCarthy and 3 others. In this proposal, McCarthy’s goal was to discover “how to make machines use language, form abstractions and concepts, solve kinds of problems now reserved for humans, and improve themselves.”
Four years later, a computer engineer at IBM by the name of Arthur Samuel coined the term “machine learning,” when he developed its first ever system, Samuel’s Checkers Player. This system was fed information from a checkers guide-book that had examples of good and bad moves in various situations. They programmed the system to learn these moves and always try to make a move as close to what the checkers guide labeled as a good move.
The theories and concepts behind AI and machine learning had existed for decades, but it wasn’t until recently that the topic of deep learning emerged. Before diving into deep learning, however, it is important to look at the basics behind machine learning.
Machine learning is a field that focuses on developing methods to give a computer the ability to learn without being explicitly programmed. In reference to Samuel’s Checkers Player, this would mean that instead of programming every move that the checkers player should make, the developer would program it to learn what moves it should take based on given data.
A self-learning computer system uses algorithms to sift through, analyze, and make meaning out of data. In this process, the system will recognize patterns in the data and construct models that allow it to form predictions and make decisions.
There are 4 broad categories within machine learning — supervised, unsupervised, semi-supervised, and reinforcement learning. Each of these categories includes their own set of algorithms, all of which relate to a specific method that the computer system takes to learn information.
Supervised, Unsupervised, and Semi-Supervised Learning
Supervised learning is a process where a human feeds a computer “training” data. This data includes labeled inputs and their corresponding outputs. With organized and labeled data, it is much easier for an AI to derive meaning.
On the other hand, the process of unsupervised learning involves feeding the AI unlabeled data. The algorithms used in these cases are perfect for when the human expert does not know what to look for. Semi-supervised learning combines the two and uses some labeled data and some unlabeled data.
Lastly, reinforcement learning is when a system learns what it should do based on the interactions it has with its environment. A reinforcement learning algorithm (called an agent) constantly iterates to discover what actions will maximize their rewards while minimizing their risks. These algorithms allow machines and software agents to maximize their performance in any given environment.
So then, how and where does deep learning fit into all of this?
Just like machine learning, the goal of deep learning is to allow an AI to make accurate predictions and decisions about new data that it is receiving. With deep learning, the difference is that it applies unique methods to allow the AI to formulate these predictions.
Deep learning takes its inspiration from the structure of the human brain and implements deep neural networks to ask a series of binary true/false questions, which help classify data according to the answers received. These networks extract features that are fed to other algorithms for clustering and classification. Think of deep neural networks as a sort of machine sensory system.
Imagine a system designed to understand what type of food it is looking at. The picture of the food would be analyzed in layers. Each layer asks a simple question about one feature of the image. Every subsequent layer adds more information for the AI to understand what it is processing.
How Is It Used?
Used in everything from voice and image recognition to the delivery of advertisements and suggestions, deep learning is at the forefront of the modern application of machine learning. Deep learning is also particularly good a feature detection.
Feature detection is the ability for a computer system to detect whether or not a type of feature exists at any given point on an image. A good example of its implementation is found in self-driving cars.
Self-driving cars use cameras and sensors to give them the ability to function effectively. Feature detection and deep learning play a big role in the car’s ability to detect objects and make decisions.
They Are More Related Than Different
When trying to understand machine and deep learning, it may be more appropriate to ask what the similarities are, rather than the differences. Being a subset of machine learning, deep learning shares many traits with its parent.
There is, however, a unique feature about deep learning that is not found in other areas of machine learning — the deep neural network. It just so happens that deep neural networks are very effective in a variety of modern applications.
At the end of the day, understanding the basics behind AI, machine learning, and deep learning will give you a better idea of where most of today’s products and services are heading. This information should give you a helpful frame of reference for thinking about how these concepts relate to each other.