Where do algorithms end and Machine Learning and Artificial Intelligence begin? In this chapter, we’ll go beyond the buzz of trendy acronyms to understand the terms, the technology behind them, and how they fit together.
What’s in a Name?
Machine Learning and Artificial Intelligence: the terms are thrown around, sometimes without much thought. After all, not just any algorithm qualifies for the label of “Machine Learning” or “Artificial Intelligence.” We talked in the last chapter about how ML is an advanced algorithm (or model)—one that learns patterns in data and then calculates similar patterns in new data. But let’s take a step back and define “algorithm.”
Simply put, an algorithm is any step-by-step process with a defined output. So yes, adding water, coffee grounds, and pressing the button on your coffee maker counts as an algorithm. And if you set the coffee maker to start percolating at a certain time, well then, you’ve got an automated algorithm. Meanwhile, data science uses analytics to extract meaning from data. At its simplest, this starts as creating a business report and builds its way to Big Data Business Intelligence. Advanced analytics turns into ML and before you know it, we’re into the broader AI field.
From Algorithms to Artificial Intelligence
At what point do we shift from simple algorithms to artificial intelligence? Some would say it’s all in how the algorithm is programmed. As the complexity of the algorithm increases, it moves closer to the AI mark.
In between the simplest algorithm and the most intricate algorithm is what is sometimes referred to as “Algorithmic Intelligence.” Algorithmic intelligence simply describes a process that can be defined as either basic or complex but with defined outputs.
Artificial Intelligence is the umbrella term covering the point at which computer processes can mimic human functions and outputs are not predefined. There are many subsets of AI, but perhaps the most well known include the following:
ML refers to algorithms taking in data and performing calculations to find an answer. It includes software code that detects patterns in data.
For example, a data scientist can use an algorithm to build (or train) a model that predicts outcomes. These models perform a range of different tasks on data. As noted in the previous chapter, ML is great for processes that search, detect, or forecast based on large amounts of data.
Another branch of AI applications includes language-related algorithms, including speech-to-text and text-to-speech applications. Common applications of this type of AI include voice assistants like Siri and Alexa, chatbots like the one in your banking app or e-commerce site, and translation engines such as Google Translate and Microsoft Translator.
Vision-focused algorithms process images. Computer vision and machine vision are often used in scenarios where large volumes of images need to be sorted or checked for errors, such as for goods packaging, braking systems for vehicles, and photo ID verification. Machine vision uses a camera to take a processable image and then partners with computer vision algorithms to process the image.
Deep Learning (DL) can be considered a follow-on to ML. Traditional ML algorithms are linear, while DL algorithms are stacked in a hierarchy of increasing complexity and abstraction. Advanced solutions like driverless vehicles and medical diagnostic programs rely on DL.
Natural Language Processing: Natural language processing (NLP) is a more complex speech-related AI. NLP combines computational linguistics—rule-based modeling of human language—with statistical, machine learning, and deep learning models. These technologies work together to enable computers to process human language as text or voice data and to comprehend its full meaning, complete with sentiment and intent.
AI Model Supervision Types
Algorithms can be simple (defined input leads to a defined output) or complex (a system comes to a defined output by using a set of complex rules, calculations, or problem-solving operations). To reach the status of AI, outputs are not defined; instead, they are based on complex mapping of user data that is then multiplied with each output.
AI models also use the concept of supervision with categories including supervised, unsupervised, semi-supervised, or reinforced learning. Supervised learning relies on a full set of labeled data when training an algorithm. “Fully labeled” means that each item in the training dataset is tagged with the answer that the algorithm should deliver. For instance, a labeled dataset of cat images would indicate which photos were of Siamese, Bengal, and Persian cats. The model then compares any new image to the training examples to predict the correct label.
Supervised learning is best for situations that include many verified data points—very usable data. Unfortunately, that is sometimes difficult to provide. That’s where unsupervised learning enters. Unsupervised learning is a deep learning model that uses a dataset without explicit instructions on what to do with it. The training dataset is a collection of examples without a specific desired outcome or correct answer. The neural network then attempts to automatically find structure in the data by extracting useful features and analyzing its structure. There are a few ways to employ this type of learning:
By looking at two or more key attributes of a data point, an unsupervised learning model can offer a prediction about the other associated attributes. For instance, if you are shopping online and look at mixing bowls, baking sheets, and wire whips, the site might suggest that you add a cake pan and parchment paper to your order.
Unsupervised learning can flag unusual items in a dataset. This can be used in defect detection in manufacturing, fraud detection in financial transactions, or spotting potential risk or medical problems in health data.
The most common application for unsupervised learning, clustering is a deep learning model that searches for training data that are similar to each other and groups them together. For instance, such a model could be used to sort photos of dogs roughly by breeds, relying on cues like the overall size, fur length and color, and head shape.
Autoencoders compress input data into a code and then try to recreate the input data from that summarized code. One application of this type of learning is improving picture quality for video or medical scans. By training with both noisy and clean versions of an image, autoencoders can remove noise and gain clarity.
You can also use semi-supervised learning and reinforcement learning. As the name implies, semi-supervised learning tries to use the best of both unsupervised and supervised models. It uses a small proportion of label data and offers improved accuracy compared to a fully unsupervised model. Reinforcement learning is a model that learns through trial and error. An example of this is the Netflix recommendation engine—it learns by making suggestions, which you take or don’t (or start watching but don’t finish), and the model responds accordingly.
As we noted in the previous chapter, ML can be especially helpful to ITOps organizations. Using a variety of model types and construction, ITOps teams can use AI and ML to detect anomalies, predict usage, and suggest solutions to IT issues. In our next chapter, we’ll look at the future of ML for ITOps, including industry trends and use-case examples.