AI & IoT Glossary

Strong artificial intelligence (AI) is a theoretical form of AI that describes a particular AI development mindset. Strong AI requires machines to be as intelligent as humans. They should thus be self-aware or able to independently solve problems, learn, and plan.

Strong AI is also known as “artificial general intelligence (AGI)” or “general AI.”

Adversarial search is a method applied to a situation where you are planning while another actor prepares against you. Your plans, therefore, could be affected by your opponent’s actions. The term “search” in the context of adversarial search refers to games, at least in the realm of artificial intelligence (AI).

Ambient intelligence, often shortened to “AmI,” is an emerging technology that aims to bring pervasive computing, artificial intelligence (AI), sensors, and sensor networks to our everyday lives. This technology is human-centric, as it is highly responsive to the presence of humans within its environment.

Ambient intelligence is a sophisticated AI system that detects and reacts to human presence. One concrete example of ambient intelligence is Siri or Alexa that responds to you once it detects your voice.

Applied machine learning refers to the application of machine learning (ML) to address different data-related problems. Its connotation is similar to applied mathematics—pure math involves many theories, which are applied and put to practical use in applied mathematics. As a result, applied mathematics helps solve real-world problems in engineering, biology, business, and many other fields.

Similarly, applied machine learning takes one's understanding of ML concepts and theories and uses these to solve problems. For example, a fundamental concept in ML is supervised learning, an ML task that requires labeled data and follows a path toward obtaining the desired output. The concept is pretty straightforward, but how can this address real-world problems?

This ML concept can be applied in computing credit scores and automating credit approval. You feed the machine with credit-related data, such as credit history and limit utilization (i.e., labeled data), and teach it to compute the credit score of a particular person (i.e., desired output). If the credit score reaches the minimum acceptable level, the credit application is approved.

Augmented intelligence is a partnership model between people and artificial intelligence (AI) systems that aims to improve cognitive performance and decision making and come up with new learning implementations.

Augmented intelligence is also known as “intelligence amplification” since its primary goal is to enhance people’s knowledge and skills by working with machines. In essence, it aims to elevate human intelligence so the user can complete tasks better and smarter.

A behavior tree is a mathematical model that details the plan of execution of a predefined set of tasks. If you have a finite set of tasks, then a behavior tree is where you specify how the program switches from one task to another.

Behavior trees were initially used to develop video games but are now widely used in robotics, computer science, and artificial intelligence (AI). With behavior trees, complex tasks, such as catching Pacman, are made possible through a series of simple tasks. In “Pacman,” these tasks include determining the locations of pills and fruits.

Cloud robotics harnesses the power of the cloud (e.g., cloud computing, cloud storage, and other cloud-based technologies) for robotics. Cloud-connected robots can use the powerful computation, storage, and communication resources of data centers to process and share information with other robots, machines, smart objects, humans, and so on. Humans can also give them tasks remotely.

Apart from faster processing, cloud robotics also allow users to reduce costs. As such, cloud robotics makes it possible to build lightweight, low-cost, smarter robots with an intelligent cloud-hosted “brain.”

Cognitive automation occurs when a piece of software brings intelligence to information-intensive processes. It has to do with robotic process automation (RPA) and fuses artificial intelligence (AI) and cognitive computing. Using AI, the process extends and improves actions typically correlated with RPA, saving users money and satisfying customers while accurately completing complex business processes that use unstructured information.

RPA is a means to automate business processes using AI or digital workers. Cognitive computing, meanwhile, allows these workers to process signals or inputs.

Cognitive robotic process automation or cognitive RPA is a subset of robotic process automation (RPA) that uses artificial intelligence (AI) technologies, including machine learning (ML), optical character recognition (OCR), and text analytics, to automate work processes.

In essence, it is a highly advanced form of RPA wherein robotic or automated processes highly mimic human activities. Most of the functions carried out by cognitive RPA systems focus on learning (gathering information), reasoning (forming contextual conclusions), and self-correction (analyzing successes and failures).

Compared with traditional RPA, which requires structured data, cognitive RPA can automate processes given unstructured information, such as emails, voice recordings, letters, and scanned documents. It can process data without human intervention, removing complex but non-rule-based tasks off human workers’ hands.

Context awareness is the ability of computing devices to sense and react to their environment. A straightforward example would be how your mobile phone changes its screen orientation depending on where you tilt it. It automatically switches from portrait (taller than broader screen) to landscape (wider than taller screen) orientation if you turn the device 90 degrees to the left or right.

A more complex example would be how your mobile phone adjusts the current time and date depending on your location. If you travel from Malaysia to the U.S., for example, the time and date shown instantly change when you turn it back on once the plane lands.

A convolutional neural network (CNN) is a deep learning algorithm that employs image recognition, processing, and classification to identify objects and detect faces. It consists of neurons that receive inputs, assign importance to them, and cluster them according to similarities.

A CNN, also called a “ConvNet,” can look at an object’s surroundings to come up with accurate predictions. Instead of looking at the whole image to determine features, it will look at layers or smaller portions.

Curiosity artificial intelligence (AI) mimics the natural curiosity of humans that allows us to learn things on our own. The goal is to develop curiosity through machine learning (ML) algorithms so AI systems can seek solutions to new problems independently.

For instance, robots can be programmed to explore their environment and learn without being told precisely what to do. This AI approach can help machines learn faster. While most ML approaches are too dependent on human intervention and instruction, curiosity AI trains systems to investigate unfamiliar data or events.

Curiosity artificial intelligence is also known as “curiosity AI,” “curious algorithm,” “algorithm curiosity,” and “artificial curiosity.”

Decision intelligence (DI) combines data science with different scientific theories to help people make the best possible decisions. It aims to provide actionable insights by translating raw data into formats that decision-makers can easily understand.

DI aims to identify trends and predict the outcomes of different decision options. A typical example of DI is a recommendation engine. When you go to Amazon, you will likely see a list of suggested products the vendor derived after analyzing your past purchases, viewed products, and search history.

An edge device is a type of hardware located at the periphery or “edge” of a network. It can perform specific tasks, such as data processing, routing, storage, filtering, and communication. These devices are typically near the data source.

Examples of edge devices include routers, gateways, sensors, and cameras. They are used in Internet of Things (IoT) applications, where they collect and transmit data to central servers or cloud-based services for further analysis and processing.

Electric field sensing refers to using a sensory system that utilizes an electric field that detects nearby objects, provided they are at least slightly conductive. One such sensory system is the People Detector, a device that senses the presence of moving and stationary objects near solid materials. It detects changes ranging between a few centimetres to 4 meters.

Electric field sensing is one of the technologies used by smart cars to detect roads and nearby cars, allowing them to avoid collisions and straying from the right path.

Few-shot learning is a machine learning (ML) technique that enables systems to learn a task or concept through examples. The technique reduces the amount of data needed to train an ML model. Instead of introducing vast volumes of training data, few-shot learning requires only a few examples for the model to learn.

Traditional ML requires you to input thousands of grammar rules when teaching a system to recognize grammatical errors. But with few-shot learning, you could train the model using only a handful of sentences with grammatical errors as examples. It will then apply what it learned from those examples to recognize grammatical errors in sentences it has never seen before. 

Forward chaining begins by inferring a set of rules or known data and going “forward” to achieve a goal. As such, it simplifies a complex task by dividing it into several simpler tasks that a computer may carry out either synchronously or sequentially, much like in a chain of processes. The data-driven and logical process of forward chaining is thus commonly implemented in production rule and expert systems.

In artificial intelligence (AI), forward chaining helps a program come up with a solution by analyzing known data and aligning it with predetermined parameters. An example would be when an end-user uses an app to determine what kind of insect he/she is looking at. The app begins by determining how many legs the insect has, what its color is, and so on until it gains enough inputs to come up with an answer.

Granular computing or GrC is an emerging model of information processing where data undergoes division into information granules. An information granule refers to a collection of entities. Entities, meanwhile, refer to the results of numerical analysis and data arrangement, depending on their functional or physical characteristics, similarities, and differences.

Granular computing is not an algorithm per se but an approach that divides information into smaller pieces to see if they differ on a granular level. The relationships seen are then used to design machine learning (ML) and reasoning systems.

A graph neural network is a deep learning method that analyzes and makes predictions based on data described by a graph. Deep learning is a machine learning (ML) technique that teaches computers to mimic the workings of a human brain. A graph in computer science, meanwhile, is a data structure that has two parts—vertices and edges. We’ll discuss these components in greater detail in another section.

Graphs in graph neural networks, however, differ from those we see and learn about in math classes. They aren’t limited to pie charts, bar graphs, and others like them. You’ll see why later on.

Independent component analysis is a computational method that separates a multivariate signal into its additive components. Sounds complicated, doesn’t it? Let’s break the definition down, then.

In computing, a multivariate signal is simply a signal that contains several distinguishable components. So you can think of it as a complete song—with music, lyrics, and even additional sound effects and backup vocals. Suppose you break the piece down into its parts. In that case, you can apply independent component analysis, so to speak, by separating each instrument, singer, and object making the different sounds from one another. In this example, the song is the multivariate signal while the instruments, singers, and objects are its additive components.

Intelligent character recognition (ICR) is an extension of optical character recognition (OCR). While OCR identifies hand-printed characters, ICR lets computers recognize different font styles to improve their text recognition accuracy.

Both technologies (OCR and ICR), in the simplest terms, allow digital devices to read text.

The Internet of Behavior (IoB) simply refers to the process of collecting all kinds of data (business intelligence [BI], big data, etc.) that shows essential information on clients’ behaviors, interests, and preferences.

The IoB allows users to understand the data from people’s online activities. It aims to interpret data and use the knowledge gained to create and promote new products and services based on human psychology.

A knowledge-based system (KBS) is an artificial intelligence (AI)-based one that uses information from various sources to generate new knowledge to help people make decisions. These devices have built-in problem-solving capabilities and rely extensively on data to provide accurate results.

KBSs are made up of two critical components—a knowledge base and an inference engine. The knowledge base contains all necessary data, while the inference engine tells the system how to process data. Most KBSs have user interfaces (UIs) to make it easy for users to send requests and interact with them.

Machine teaching is a subfield of artificial intelligence (AI) that pertains to the process of obtaining knowledge directly from people instead of merely extracting knowledge from data. The idea is to provide contextualized data to AI systems to bring outputs relevant to their users.

Machine teaching can be considered the reverse of machine learning (ML). In machine teaching, the system acts as a teacher who begins with a goal in mind rather than a desired result. The teacher develops an optimal training process that allows the learner to achieve that goal. In short, the teacher makes it easier for the learner to process and overcome problems. Ideally, machine teaching involves deconstructing the problem into smaller parts that are easier to solve for ML algorithms.

Machine teaching provides immense benefits in supervised learning scenarios where ML algorithms have little or no labeled training data to produce specific outcomes.

Machine-to-machine authentication refers to the process of allowing different remote systems to communicate with each other. Your favorite vending machine, for example, can be set up to automatically send an order to the supplier’s system for items that are running out of stock.

Since machine-to-machine communication can occur over wired, wireless, and any form of channel, it is prone to abuse and glitches. Machine-to-machine authentication helps ensure that only authorized services can access information on another system. It wouldn’t be ideal for a threat actor’s or competitor’s application to access the vending machine supplier’s inventory and order filing system.

Most machine-to-machine authentication solution providers use Open Authorization 2.0 (OAuth) as a way for applications to access systems. OAuth 2.0 is the current standard protocol for online authorization. It replaced the first version, OAuth.

Model-based reinforcement learning is a means for machines to make decisions using a predictive model to determine what will happen if a particular course of action is taken to choose the best solution. Let’s break the definition down to understand the concept better.

A predictive model is used in statistics to predict outcomes by analyzing patterns in a data set. Think of it as a list of all possible outcomes from which a machine will choose the one that best suits the problem presented to it.

Model-free reinforcement learning is a reinforcement learning algorithm that does not use the transition probability distribution and the reward function of the Markov decision process that represents the problem that requires solving.

In model-free reinforcement learning, the transition probability distribution or transition model and the reward function are collectively called the “model.” Their absence in the process is why the name “model-free.”

Since the reinforcement learning algorithm does not use a model, you can think of it as a trial-and-error algorithm. The machine tries all possible solutions until it gets the best result.

The Massachusetts Institute of Technology (MIT)’s Moral Machine is a platform that gathers human perspectives on moral decisions made by machine intelligence (e.g., self-driving cars). The system generates moral dilemmas and lets a driverless car choose the lesser of two evils (e.g., Should it kill two passengers or five pedestrians?). Outside observers (i.e., people) judge which outcome they think is more acceptable. They can see how their responses compare with those of others afterward. The platform also lets people create their own scenarios for others to view, share, and discuss.

Natural language generation (NLG) is a type of artificial intelligence (AI) that generates natural language from structured data. While that sounds complicated, it simply means translating massive amounts of information into something humans can read and understand. But the data needs to be appropriately formatted for NLG to work.

An example of NLG would be translating the numbers in a spreadsheet into narratives or words to create human-readable text. NLG uses machine learning (ML), deep learning, and neural networks to make this process possible.

Neural Architecture Search (NAS) is the process of discovering the best architecture a neural network should use for a specific need. In the past, programmers had to tweak neural networks to learn what works well manually. NAS automated that process, allowing artificial intelligence (AI) systems to discover more complex architectures.

In essence, NAS uses various tools and methods that test and evaluate huge architecture volumes utilizing a search strategy to choose one that best meets the programmer’s objectives.

The noisy channel model is a framework that computers use to check spelling, answer questions, recognize speech, and perform machine translation. It aims to determine the correct word if you type its misspelled version or mispronounce it.

The noisy channel model can correct several typing mistakes, including missing letters (changing “leter” to “letter”), accidental letter additions (replacing “misstake” with “mistake”), swapped letters (changing “recieve” to “receive”), and replacing incorrect letters (replacing “fimite” with “finite”).

A perceptron is a single-layer neural network for supervised learning. A neural network is a computer that mimics the way a human brain works. Supervised learning, meanwhile, is a machine learning (ML) technique that requires a human to train a computer by giving it data to get the desired results.

A perceptron has four main parts that we will describe in greater detail in the following section. These are input values, weights and bias, the net sum, and an activation function.

Predictive modeling is a mathematical technique that uses historical data and results to create a model that can predict future outcomes. For example, banks use predictive modeling in credit scoring to foresee a potential client’s ability to repay a loan. Airlines can also use the technique to predict the volume of passengers for a particular month or season.

Some may argue that predictive modeling rests on the premise that history repeats itself, which is not too far-fetched. After all, financial predictive models can predict if a client is likely to make late payments in the future based on his or her past behavior.

Prompt-based learning is a machine learning (ML) strategy that uses pretrained language models to train large language models (LLMs) so that the same model can be used for different tasks without retraining. It utilizes the knowledge acquired by the pretrained language models on a large amount of text data to solve various downstream tasks, such as text classification, machine translation, named entity detection, text summarization, and others.

Also known as “prompt learning,” prompt-based learning is an emerging strategy to allow pretrained artificial intelligence (AI) or foundation models to be repurposed for other uses without additional training. It is a new natural language processing (NLP) paradigm that does not require any supervised learning process since it directly relies on the objective function of any pretrained language model. It is a simple implementation of using a language model for a prompt-based learning model.

A rational agent is an artificial intelligence (AI) component. It applies AI to different real-world problems. As such, it chooses an action from a set of distinct options. It has models that allow it to deal with unexpected variables and always selects the best possible outcome from all the available options.

The term “rational agent,” however, is not only applied to a system. It can also refer to a person, a company, or an application, practically anything or anyone that makes rational decisions.

Recursive Bayesian estimation is an approach used in statistics and machine learning (ML) that estimates the current state of a system. The framework is used in robotics and automotive technology, where machines are taught to perform a task that requires estimation.

A self-driving car, for example, can estimate its location using the recursive Bayesian estimation framework. The car obtains its starting position via a Global Positioning System (GPS). Its built-in algorithms then help it estimate its current location after a certain amount of time or distance. The algorithms often use the Bayesian mathematical concept in statistics.

A robotic process automation (RPA) bot is a software robot or application that automates repetitive and rule-based tasks within business processes. Its design lets it mimic human interactions with various software, such as entering data, manipulating applications, triggering responses, and communicating with other systems.

RPA bots typically operate at the user interface (UI) level, interacting with applications and systems like human users. They can navigate through different screens, extract data from documents, perform calculations, make decisions based on predefined rules, and even communicate with other bots or systems through application programming interfaces (APIs).

Self-supervised learning is a means for training computers to do tasks without humans providing labeled data (i.e., a picture of a dog accompanied by the label “dog”). It is a subset of unsupervised learning where outputs or goals are derived by machines that label, categorize, and analyze information on their own then draw conclusions based on connections and correlations. 

Self-supervised learning can also be an autonomous form of supervised learning because it does not require human input in the form of data labeling. In contrast to unsupervised learning, self-supervised learning does not focus on clustering and grouping that is commonly associated with unsupervised learning. 

Spatial computing refers to the process of using digital technology to make computers interact seamlessly in a three-dimensional world using augmented reality (AR), virtual reality (VR), and mixed reality (MR). Spatial computing uses physical space to send input and receive output from a computer.

Through spatial computing, most people no longer interact with computers as an outsider would. Instead, they get to experience what it’s like to be within the digital realm by interacting with objects that only exist in it. The concept allows the marriage of the real world and the digital landscape.

A support vector machine (SVM) is a machine learning (ML) algorithm that employs supervised learning models to solve complex classification, regression, and outlier detection problems. It performs optimal data transformations that determine the boundaries between data points based on predefined classes, labels, or outputs.

Since it requires supervised learning, teaching it to do complicated tasks needs human input. Think of it as learning with the help of a teacher instead of studying independently.

The Theory of Computation is a branch of computer science and mathematics that focuses on determining problems that can be solved mechanically, using an algorithm or a set of programming rules. It is also concerned with the efficiency at which the algorithm can perform the solution.

In simple terms, the Theory of Computation answers these questions:

• What problems can the machine solve? What problems can’t it solve?
• How fast can a machine solve a problem?
• How much memory space does a machine need to solve a problem?

To answer these questions, computer scientists use a model of computation, which is a computer simulation for the algorithm being developed. The Turing machine is among the most used models of computation.

Virtual intelligence is a type of artificial intelligence (AI) that exists inside a virtual world. It is created and optimized to carry out specific tasks to aid a user within a defined framework. At the outset, a virtual intelligence machine can seem like a smart system based on how it interacts with a user. In reality, though, it isn’t.

Virtual intelligence is a code or program that functions within the controlled environment it was created for. As such, it can’t generate spontaneous solutions and responses. It responds only based on predetermined factors.

You can find virtual intelligence in action in Global Positioning System (GPS) navigation software, chatbots, virtual assistants, interactive maps, and wearables. While it can learn from its interactions to enhance performance, that learning is limited to the original functionality it was designed for.

A Vision Transformer (ViT) is a deep learning model architecture that applies the transformer architecture, initially introduced for natural language processing (NLP), to computer vision tasks. It is designed to process and understand visual data like images in a way comparable to how transformers process and understand sequential data in language processing tasks.

While convolutional neural networks (CNNs) used to be the dominant architecture for image analysis and computer vision tasks like image classification, object detection, and image segmentation, they are inherently limited by their reliance on local spatial relationships and struggle to capture global context and long-range dependencies within an image. ViTs overcame these limitations by adapting the transformer architecture, which proved superior for language tasks that require long-range dependencies.