Forms of AI and approaches

Implementations of AI have come in various forms due to the varying nature of the intended application and the technology available for the solution. Hence, AI has been manifested in code in various forms, utilized by a wide range of developers in different domains for respective problems.

In the following Venn diagram, we can see various forms of AI:

In the preceding diagram, I have mentioned all the major forms of AI categorized into three major manifestations. Each form is explained in detail in the following section, broken down into expert systems, machine learning, and neural networks.

We will now explore these primary approaches and forms of AI with brief introductions to their backgrounds and applications.

Statistical and expert systems

Statistical systems were one of the most primitive forms of AI, dating back to the late 1960s. As the name suggests, statistical approaches used a huge amount of data to arrive at the most desirable result. However, it was soon recognized that the results were virtually unrelated to real-world scenarios and produced output only based on the AI’s rational decision-making ability. These limitations led to the decline of statistical AI, paving the way for expert systems in the early 1980s.

Expert systems were a mature form of strong AI with the ability to mine datasets and derive answers that were more related to the context of the problem. This leap was aided by information theory, combined with new abilities in hardware. Although expert systems were developed in the early 1960s, they only became affordable during the 1980s thanks to the PC revolution. Unlike the scientific approaches used by statistical AIs, expert systems leveraged semantic and linguistic programming to arrive at the expected outputs with high probability.

An example of an expert system in use can be seen in the following photo:

You can check out the preceding photograph by Michael L. Umbricht and Carl R. Friend at https://en.wikipedia.org/wiki/Expert_system#/media/File:Symbolics3640_Modified.JPG. Here is the license to the photo, https://creativecommons.org/licenses/by-sa/3.0/.

Although expert systems opened the doors for early AI adoption, it is machine learning that really met the demands of industry. In the following section, we will learn about machine learning.

Machine learning

Machine learning is a form of AI that depends on a preexisting dataset as input, with or without a variation in the expected output to produce human-like thinking based on applying a mathematical model on the given data. The term was coined in 1959 by Arthur Samuel, one of the pioneers of AI research at IBM. If a particular machine learning algorithmic system aims to extrapolate a result based on the given forecast data, it is called predictive analytics, which is used in various emerging fields of computer applications. 

Although similar forms of AI existed before machine learning, it is believed that most of the research has now been consolidated under this label since the early 1990s, also known as the golden age of machine learning. Some of the earliest applications of the concepts of machine learning were CV, email spam filtering, and operation optimizations.

There are three approaches to machine learning algorithms that have been observed in use very consistently in the recent past. We will look at them in the following sections.

Supervised learning

Models in this approach directly depend on the datasets that serve as the input for training data, and also on the expected outputs. The model uses the input data in the training phase by itself, learning outcomes associated with a few ranges of inputs in the form of labeled samples. Such samples are fed into the algorithm model to be able to successfully achieve the expected result. Usually, the expected outcomes are either in the form of classification, regression, or prediction.

Unsupervised learning

Under this approach, the models are provided with the training data as the input, but lack any expected output to be specified by the end user. This approach is justified as the intended outcome of this practice is to gain visibility on the unexplored rational commonalities present within the data.

Reinforcement learning

This is a reward-based learning approach used to achieve all-round optimization by enforcing a wide range of techniques in rewarding successful agents in a cumulative manner.

Now that we have a basic understanding of machine learning, let's proceed to examine neural networks.

Neural networks

An Artificial Neural Network (ANN), also called deep learning, is a group of synthetic neurons forming a circuit to solve difficult problems. This is a specialized form of AI with aggressive strategies designed to achieve the desired goal. However, unlike machine learning algorithms, the heuristics and execution patterns in ANNs are not linear, and hence this kind of AI can be found in a wide range of applications such as autonomous driving, textual and facial pattern recognition, decision-making software for trading, digital art, and drug formulation.

The following diagram is a general representation of a neural network, along with the basic relationship between the three layers—Input, Hidden, and Output:

You can check out the preceding diagram by Glosser.ca at https://commons.wikimedia.org/wiki/File:Colored_neural_network.svg. Here is the license to the diagram, https://creativecommons.org/licenses/by-sa/3.0/.

Evolutionary computation

AI has long been identified as a key enabler for the future of biotechnology. Evolutionary AI in forms such as genetic algorithms have also been one of the early fields of research in this domain. AI has been helpful in analyzing, simulating, and predicting the behavior of mutations in our bodies. It is also notable that some AI practices in genome research have been actively criticized, fearing severe repercussions for the future of mankind in the process of experimentation.

Swarm computation

Apart from behaving in a centralized manner, AI is also significantly known to have disrupted the functioning of distributed and collaborative computer systems. Swarm intelligence is the capability of a group of systems to achieve a common goal by cooperating in an ordered manner. Swarm intelligence is leveraged to understand group behaviors and optimize processes wherever possible.

Multiple agents work together based on a set of heuristics to consume vast amounts of data and produce meaningful results based on the coordination between one or more computing devices. Applications of swarm AI can be observed in robotics, logistical automation such as truck platooning, and so on.

The following photograph is a real-world example of a coordinated application using swarm computation techniques:

You can check out the preceding photo by Serge Kernbach at https://commons.wikimedia.org/wiki/File:RechargingSwarm.jpg. Here is the license to photo, https://creativecommons.org/licenses/by-sa/3.0/.

With this basic understanding of AI and its types, forms, and approaches, we will now explore the procedure of applying AI in the next section.