11 Important AI Terms and Concepts Explained With Clarity

From algorithms to recursion: A glossary of some of the essential terms in the AI world

There are many terms and concepts that can be confusing in the world of artificial intelligence. This piece can be used as a glossary of some of the most popular terms in the industry.

Algorithm:

A set of instructions designed to perform a specific task. AI relies on algorithms, but a single algorithm does not necessarily

constitute AI and machine learning.

Artificial intelligence:

Widely defined as the engineering and science of creating intelligent machines. AI uses computer programs and models based on networks of databases in a process that teaches computers. Using data and high-level mathematical functions, the computers incorporate statistical data analysis, labels, and machine learning to enable machines to learn, adapt, and perform tasks previously thought to require human cognitive functions.

AI-enabled:

A machine that was designed and developed using AI but that does not continue learning (machine learning) once operational. The FDA calls this type of AI software locked.

Bias:

A systematic deviation from the truth (vs variance, which is random deviation). Bias in the data training set will affect the results

of algorithms that are run for machine learning, regardless of the algorithms’ accuracy. Some bias is inherent in most training data sets, although rare as result. Although bias occurs, it most often occurs unintentionally. In algorithmic bias, a software system reflects the implicit values of the humans creating it or the data sets training it.

Once a machine starts learning, its calculations become more complex. The machine might make automatic modifications that lead to bias or variances. A concern about bias is not the end of the software but an opportunity for people to investigate and intervene using a structured approach to correct machine learning. Automation bias also can occur; this bias is the tendency for people to favor decisions generated by machines at the expense of ignoring conflicting data or human decisions.

Labels:

Labels add context, information, and value to image data. The labeled data represents ground truth, or prior knowledge, on

which to base machine learning. In medical imaging and radiation therapy, labels are study-level descriptors (eg, head CT or abdominal MR) or image-level descriptors (eg, pixels on image 25 represent the kidney). Label sources include image annotations, such as measurements, radiology report findings, electronic health record clinical data, and other data generated by professionals for use in AI.

Machine learning:

Use of algorithms that are designed and trained to change their performance (adapt) as it focuses on specific tasks, and ideally improve with exposure to data to achieve AI. The machines are programmed to review large data sets, look for patterns, make predictions, and act on the data. This ability makes the software running them adaptive in behavior (according to the FDA)

because the machines learn and change based on new data.

Supervised learning:

Labeled training data are input into the machine to facilitate machine learning that can be generalized based on the training data. A teacher or “critic” also can provide feedback and correction as the machine learns.

Unsupervised learning:

The machine learns based on programmed patterns with no explicit labeling or feedback from a teacher. This type of learning looks for density or clustering of data but learns only the inherent structure of that data.

Deep learning:

The basis for most image interpretation machine learning that most closely resembles learning by the human brain. The computer learns based on multiple layers of interconnected algorithms categorized into a hierarchy of importance. Deep learning can be supervised, unsupervised, or partially supervised. It is responsible for providing practical applications of machine learning.

Neural network:

A neural network is a layered set of algorithms that is modeled to resemble the human brain. The neural network looks for patterns through labeling or clustering of raw data.

Recursion:

One of 2 common ways of solving complex problems. In iteration, a problem is converted into a series of steps that occurs in a set order. In recursion, the steps are piled on one another and replicate themselves at smaller scales until they combine to solve the

problem.

Follow me for daily news and updates from the world of AI, machine learning, and data science.