Today, the terms artificial intelligence (AI) and machine learning (ML) are often used interchangeably. While the terms are related, they mean different things. Disentangling the complicated relationships between these terms can be a difficult task. We’ll try to map out how they all relate to one another, so your team can find the best candidates, best approaches and best frameworks as you embark upon your AI journey.
Artificial Intelligence (AI) Is a State of Being
AI refers to the concept of machines mimicking human cognition. The original description from the Dartmouth Summer Research Project proposal in 1956 describes AI as:
“the conjecture that every aspect of learning or any other feature of intelligence can in principle be so precisely described that a machine can be made to simulate it.”
In practice today, we see AI in image classification for platforms like Pinterest, IBM’s Watson picking Jeopardy! answers, Deep Blue beating a chess champion, and voice assistants like Siri responding to human language commands. To reference artificial intelligence is to allude to machines performing tasks that only seemed plausible with human thinking and logic.
In the real world, one of the most ubiquitous forms of AI might manifest themselves in the form of conversational AI. Conversation AI may include multimodal inputs (e.g. voice, facial recognition) with multimodal outputs (e.g image, synthesized voice). All of these modalities can be considered part of AI, as well as the integration of these modalities.
You may already be familiar with some of these modalities. A few years ago, Starbucks enhanced its mobile app by enabling ordering ahead via voice commands. Staples’ Easy System allows customers to order via voice commands. The National Hockey League rolled out a chatbot for easier communication with fans earlier this year. These applications of AI are examples of machines understanding human intents and returning relevant results.
Machine Learning Algorithms Create AI
Machine learning, deep learning, and active learning, on the other hand, are approaches used to achieve AI. The classical definition of machine learning (ML), like AI, dates back to the 1950s. In his paper about using machine learning to teach a computer the game of checkers, Arthur Samuel introduces his research about the “programming of a digital computer to behave in a way which, if done by human beings or animals, would be described as involving the process of learning.” Eventually, the goal of ML is that “Programming computers to learn from experience should eventually eliminate the need for much of this detailed programming effort.”
If AI is when a computer can carry out a set of tasks based on instruction, ML is a machine’s ability to ingest, parse, and learn from that data itself in order to become more accurate or precise about accomplishing that task. While other statistical methods for learning exist, through recent ML advancements, practitioners have revived the concept of neural networks, which are a series of algorithms that act—as one might assume—like the human brain.
Machine learning is prevalent anywhere AI exists, but it has some specific use cases with which we may already be familiar. Companies like Microsoft leverage predictive machine learning models to make better financial forecasts. These models make predictions on financial entities by learning from historical trends and generating forecasts of a stock’s movement.
Professional sports teams use machine learning to better project prospects during entry drafts and player transactions (trades and free agent signings). By feeding years of historical prospect data into machine learning algorithms, for example, draft teams can more accurately assess what types of statistical profiles are likely to lead to (good) professional players. In this application, algorithms learn how to better identify potential star players and, ideally, avoid draft busts.
Deep Learning, Weights and Neural Network Activity
As machine learning has advanced, researchers and programmers have dived deeper into what algorithms are able to accomplish. A layer beyond machine learning, we find deep learning. There are a few, similar definitions of deep learning. The simplest definition for deep learning is that it is “a set of algorithms in machine learning that attempt to learn in multiple levels,” where the lower-level concepts help define different higher-level concepts.
Within a neural network, each processor or “neuron,” is typically activated through sensing something about its environment, from a previously activated neuron, or by triggering an event to impact its environment. The goal of these activations is to make the network—which is a group of machine learning algorithms—achieve a certain outcome. Deep learning is about “accurately assigning credit across many such stages” of activation.
Google Brain may be the most prominent example of deep learning in action. Researchers presented to their neural network 10 million images of cats taken from YouTube videos without specifying any parameters for cat identification. The network successfully identified cat images without using labeled data.
In the real world, deep learning can also be used to synthesize new data instances, which is something traditional ML cannot do so well. One example of this in the real world is in medicine. Researchers at UCLA leveraged deep learning to enhance microscopy practices. With deep learning, researchers implemented a “framework [that] takes images from a simple, inexpensive microscope and produces images that mimic those from more advanced and expensive ones.”
Deep learning also often appears in the context of facial recognition software, a more comprehensible example for those of us without a research background. The face ID on iPhones uses a deep neural network to help phones recognize human facial features.
Active Learning Chooses Its Own Data
Most ML algorithms require annotated text, images, speech, audio or video data. But, with the right resources and the right amount of data, practitioners can leverage active learning. Active learning is the philosophy that “a machine learning algorithm can achieve greater accuracy with fewer training labels if it is allowed to choose the data from which it learns.” In order to choose the data from which it learns, an active learning-based AI can ask queries of humans in order to obtain more data.
In situations where data is not readily available or and providing labels for that data is difficult, active learning poses a helpful solution. Luckily, active learners can learn to label data themselves. If presented with a set of labeled data, active learning algorithms can ask human annotators to provide labels to unlabeled pieces of data. As humans label data, the algorithm learns what it should ask the human annotator next.
It is difficult to pinpoint specific examples of active learning in the real world. This is a difficult task in part because active learning is better thought of as a method of training machine learning algorithms, which means the technique may or may not be used in instances where machine learning drives artificial intelligence. In practice, the idea behind active learning is that data scientists can use poorly trained AI to help identify — through a Query Strategy, as outlined above — which pieces of data should be used to train a better version of that AI.
Human labelers are required for any sort of ML, but with Active Learning their work is significantly reduced by the machine selecting the most relevant data.
Differences Abound Despite Inextricable Links
Deep learning is a more advanced form of machine learning, which is used to create artificial intelligence. Active learning leverages readily available, and often imperfect, AI to actively select new data that it believes would be most beneficial when developing the next, improved version of the AI. Active Learning, therefore, can significantly reduce the amount of data required to develop a performant AI system because it only learns from the most relevant data. All of these terms are interconnected, but each refers to a specific component of creating AI. With the right understanding of what each of these phrases entails, you can get off on the right foot creating our own AI.