The B2B and B2C markets are abuzz with the terms artificial intelligence, machine learning, and deep learning. But what do these terms mean, exactly? They’re often used very loosely and many think they’re interchangeable. But they’re not. Here’s a short overview to help you cut through the static to determine which solution might be right for you and your business.
Artificial intelligence (AI) is the term for the broad discipline that includes anything related to developing machines that are “intelligent” through programming. This includes many daily items you’re familiar with, from smartphones and marketing software to chatbots and virtual assistants.
Machine learning (ML) refers to machines and systems that can learn from “experience” supplied by data and algorithms. ML is often used interchangeably with AI, but it’s not the same thing — ML is a developmental outgrowth of AI.
Deep learning (DL) is a further developmental outgrowth of ML, but applied to even larger data sets. It uses multi-layered artificial neural networks to deliver high accuracy in assigned tasks.
In terms of historical development, AI came first. It serves as the foundational discipline from which ML evolved. And ML is the foundational discipline from which DL evolved. One way to conceptualize their relationship to one another is as nested arenas of AI development along a timeline:
Artificial intelligence and how it works
In its broadest sense, AI refers to machines programmed to act according to well-defined rules and responses. The responses are confined to the set of rules that are provided, and the machines can’t deviate from those rules, except if they fail.
A very basic example of AI would be your clothes dryer. You can set a specific time and temperature and the machine performs the task according to the instructions given. It doesn’t have the ability to make decisions or make any changes by itself. A more sophisticated example would be configuring your CMS to deliver personalized website experiences. By analyzing a targeted selection of data points about your customer and writing the appropriate logic, your website can display the most relevant content. In neither case is the machine capable of being more than its programming — even if that programming makes the machine very capable in accomplishing its assigned tasks.
Machine learning and how it works
“ML is the science of getting computers to act without being explicitly programmed.” –Stanford University
Machine learning is a different approach to developing artificial intelligence. Instead of hand-coding a specific set of rules to accomplish a particular task, in ML the machine is “trained” using large amounts of data and algorithms that give it the ability to learn how to perform a task.
Over the years, the algorithmic approaches within ML have included and evolved from decision tree learning, inductive logic programming, linear or logistic regressions, clustering, reinforcement learning, and Bayesian networks. Currently, there are three general models of learning used in machine learning:
At this point in time, most machine learning is supervised, which still requires a lot of human intervention to accomplish the training. For example, a “supervisor” has to manually tell the spam filter what to look for in spam vs. non-spam messages (e.g. look for the words “Western Union” or look for links to suspicious websites, etc.) until the machine has gained enough “experience” to learn and accurately apply the distinctions.
The training goes something like this. The algorithm would be first trained with an available input data set of millions of emails that is already tagged with the spam/not spam classifications to train the ML system on the characteristics or parameters of the ‘spam’ email and distinguish it from those of ‘not spam’ emails.
In general, unsupervised learning is more difficult to implement than supervised learning. But it’s very useful when your example data set has no known answers and you’re searching for hidden patterns. The system has to train itself from the data set provided. Two popular types of unsupervised learning are clustering and association.
Clustering groups similar things together and consists of dividing a set of elements of the existing data set into groups according to a heretofore unknown pattern. For example, defining a customer demographic and further clustering them based on education or income that might affect their purchasing decisions in favor of one product or another. This would allow you to specifically target each cluster of customers more effectively.
Association involves uncovering the exact rules that will describe the larger portions of your data. For instance, “People who buy X also tend to buy Y.” For example, online book or movie recommendations are based on the association rules uncovered from your previous purchases or searches. Association algorithms are also used for purchasing-cart analysis. Given enough carts, the association technique can help predict another item you might like to put into your cart.
The machine learning system learns by trial and error through training characterized by receiving virtual rewards or punishments. Reinforcement learning comes from child-development research that says instead of telling a child which piece of clothing to put into which drawer, you reward a child with a smile when the child makes the right choice on their own or you make a sad face when the child makes the wrong choice. After just a few iterations a child learns which clothes need to go into which drawer.
In reinforcement learning, the very complex algorithms are designed so that the machine tries to find an optimal solution. It operates according to the principle of reward and punishment, and by this approach it moves quickly through several mistakes or near mistakes to the correct result by adjusting the weight of the previous results against the desired outcome. This allows the machine to make a different, better decision each time until it is rewarded.
Deep learning and how it works
Deep learning is the newest area of ML and AI that uses multi-layered artificial neural networks to accomplish tasks such as object detection, speech recognition, and language translation − all with an extremely high degree of accuracy. The artificial neural networks (ANN) are inspired by the biology of the human brain, specifically the organic interconnections between neurons.
The human brain analyzes information it receives and identifies it via neuron connections according to past information it has stored in memory. The brain does this by labeling and assigning information to various groups, and it does this in nanoseconds. Similarly, when a system receives an input, the deep learning algorithms train the artificial neurons to identify patterns and classify information to produce the desired output. But, unlike the human brain, artificial neural networks operate via discrete layers, connections, and directions of data propagation.
Despite the level of sophistication of its algorithms, DL is still just another method of statistical learning that extracts features or attributes from raw data sets. The major difference between deep learning and machine learning is that in the latter you need to provide the features manually. DL algorithms, on the other hand, automatically extract features for classification. This ability requires a huge amount of data to train the algorithms. The accuracy of the output depends on the amount of data, and deep learning requires huge data sets.
Additionally, due to the sophisticated algorithms, deep learning requires very powerful computational resources. These are specially designed, usually cloud-based computers with high performance CPUs or GPUs.
There are several kinds of artificial neural networks and DL processing applications you may have already heard of:
Convolutional neural networks (CNN) are deep artificial neural networks that are used to classify images, cluster them by similarity, and perform object recognition. These are algorithms that can identify faces, tumors, and navigate self-driving cars.
Generative adversarial networks (GAN) are composed of two neural networks: a generative network and a discriminative network. GANs are very popular in social media. If you feed the GAN with a large enough data set of faces, it can create completely new faces that are very realistic but nevertheless fake.
Natural language processing (NLP) is the ability to analyze, understand, and generate human language, whether text or speech. Alexa, Siri, Cortana, and Google Assistant all use NLP engines.
Putting IA, ML, and DL to work for you
What most of us think of as AI is, more accurately, machine learning. But understanding the history, development, and distinctions between artificial intelligence, machine learning, and deep learning can help you determine which solution would be right for your goals. The solution you choose, however, is also dependent on the amount and type of data you have access to.
Within the last couple of years, almost every company is using machine learning or deep learning (and therefore, by definition, artificial intelligence) in some capacity to move their business forward. The competitive gauntlet has been thrown down.
Fortunately, tools that were previously only available to enterprise-size companies are now affordable and accessible to mid-market companies, making machine learning the most accessible playground right now. Fusion Alliance provides cloud infrastructure and other ML services that accelerate machine learning modeling, training, and testing to our banking, financial, and retail customers.
Read more about Fusion’s work in AI, ML, and DL: