Classification of AI

Introduction

Rapid advancements in data analytics and robotics have brought artificial intelligence to the forefront of public consciousness. Even though artificial intelligence fascinates a lot of people, terms like machine learning and deep learning remain thoroughly misunderstood.

This article, then, will go through the different types of AI, existing or hypothetical, clearly showing the differences between each type.

Difference between Artificial Intelligence and Machine Learning

Artificial intelligence and machine learning are used interchangeably by many people. Even though the two are similar in many ways, they have some noticeable differences.

Big Data and AI

Big data refers to huge and complex datasets that are used to improve AI’s analytical and decision-making capabilities. Big data is often too overwhelming to be analysed and processed by humans. The two thus have a synergistic relationship, as big data analytics are improved by AI.

AI and Machine Learning

To put it simply, machine learning is a subset of artificial intelligence.

In general, AI is when a computer system has the ability to mimic human intelligence, by performing tasks that include problem-solving and learning. 

Machine learning is an application of AI. Computers are able to use data models to learn and improve without human intervention. With machine learning, computers are able to learn from experience.
A popular machine learning tool at the moment is ChatGPT, developed by OpenAI. ChatGPT is a chatbot that can communicate with a user, to provide information or solve problems. The chatbot is able to learn from experience, improving its answers as it goes along.

History of Artificial Intelligence

Although AI seems like an extremely recent phenomenon, it has a decades-long history. Here are a few moments in the history of AI:

  • 1940: The first form of Artificial Intelligence is Alan Turing’s Enigma Machine, used to crack German code during WW2.
  • 1956: John McCarthy coins the term “Artificial Intelligence” at the first AI conference
  • 1969: Construction of Shakey, the first fully-mobile robot that is able to perform tasks with a purpose, rather than just following a list of instructions. 
  • 1997: The supercomputer Deep Blue (developed by IBM) shocks the world by beating the world chess champion. During the match, Deep Blue is seen taking time to “think” about which moves to make.
  • 2008: Development of voice recognition with iPhones and Siri.
  • 2020: GPT-3 revolutionises automated conversations.

Types of Artificial Intelligence

Artificial Intelligence can be categorised into four distinct types.

Reactive AI

This is the most basic type of AI, performing actions based on the input it receives. Reactive AI will always react to the same input in the same way, making outputs easier to predict.

Reactive AI has become very sophisticated, but remains unable to learn based on past experience. Reactive AI will thus never improve its own functioning by itself.

This type of AI has served as the foundation for developing more complex and autonomous forms of AI.

Some famous examples of reactive AI are IBM’s Deep Blue (mentioned above), search or recommendation engines, and filters that label certain emails as spam.

Limited Memory AI

Limited Memory AI builds on reactive machines by having the capability to build on past data. This AI can learn from experience and make decisions.

Most AI systems, notably deep learning programs, fall under this category. 

Limited Memory AI is used for chatbots, self-driving vehicles, and virtual assistants, among others.

This type of AI has no long-term memory, however, which limits its decision-making power. A chatbot, for example, only remembers the information it receives in a given session.

Theory of Mind AI

Theory of Mind is an innovative form of AI, on which scientists are currently working. 

The big leap with Theory of Mind AI is the development of artificial emotional intelligence, the ability of a machine to recognise, understand, and remember complex human emotions. 

Artificial emotional intelligence will greatly improve a machine’s decision-making capability and will enable computers to adjust their behaviour according to identified emotions. 

Although exciting, artificial emotional intelligence remains rather elusive to scientists, as human emotions are complex and fluid phenomena. 

Some breakthroughs, however, are Kismet, a robot head able to recognise and replicate emotional signals, and Sophia, a humanoid robot that recognises faces and has her own facial expressions in response to interactions.

Self-aware AI

This next type is, as the name suggests, AI that is aware of its own existence. Self-aware AI would be aware of its own emotions, granting it a level of consciousness and decision-making power comparable to humans.

Self-aware AI is the most advanced form of AI, but has not at all been developed by scientists as of yet. Thus, self-aware AI remains only hypothetical.

Read more here for other hypothetical forms of AI.

Different types of analytics

Across all types of AI, there are various types of analytics, which is what the machine does with the software it receives.

The most common and simple are descriptive systems, which only receive data and describe it.

Predictive systems, more complex than the previous, take data to anticipate what may happen. They function with probabilistic systems. These systems are used by industries to make predictions and gain insight.

Prescriptive systems are the most complex and are used to recommend solutions based on the data they receive. These systems are used by Internet companies. 

Machine and Deep Learning

Machine and deep learning models are predictive or prescriptive models, as they use data to learn and make decisions.

Machine learning can be supervised, which is a model to determine the relationship between a given input and a given output. 

If a human does not know how to classify the data they have, they can use unsupervised machine learning models. These models are given only inputs, and serve to find patterns in data. 

There are also reinforcement models, in which an algorithm progressively learns to perform a task and maximise the rewards it reaps from certain actions. This model is useful when there is little training data.

Deep learning has various models:

  • Convolutional neural networks (CNN): these networks extract complex data to determine outputs. CNNs are often used for image recognition models, to find and extract features from an image. 
  • Recurrent neural networks (RNN): these focus on processing sequential data, so they remember and use past data. RNNs are used for language translation and speech recognition. 
  • Transformers: these are used to understand what words mean within context. This makes transformers great tools for understanding and translating languages.
  • Generative adversarial networks (GAN): these networks produce new data similar to training models with the aim of obtaining better performing models. 

Read more here for an extensive guide to AI.

Levels of AI

Traditional Machine Learning

Some machine learning models require human intervention to function. This is the case for supervised machine learning models. When finding a relationship between given inputs and outputs, a model has to be trained.

Data has to be fitted to models, which refers to determining how suited a model is to a given dataset. By finding the best parameters for the model, it will be able to provide the most precise predictions or classifications. 

The process of finding the best model is known as approximating a function. 

Various inputs can be selected or changed in machine learning models. This is known as manual feature engineering. Inputs are changed with the aim of improving a model’s performance. 

Feature learning and federated learning

Feature learning is used in unsupervised learning models and refers to a computer’s ability to learn useful features from input data. Feature learning is used to improve a machine learning model’s performance.

With feature learning, a model can find patterns and relationships in data without the need for guidance.

Federated learning, on the other hand, enables multiple parties to work on a machine learning model without sharing raw data. There is no longer a central server that trains models; local models are trained and model updates are then sent to the central server.  

Federated learning has multiple benefits, including improved privacy, as input data is not shared and scalability — decentralised work is more conducive to developing performative large-scale models.   

This form of machine learning has been especially good for mobile platforms and the Internet of Things. 

Transfer Learning

Transfer Learning refers to the use of information gained from one model on a separate model. In other words, information is stored from the first model, and then altered to suit the next model.

Transfer Learning enables rapid improvement of models, and resembles reinforcement learning, but across multiple models. 

Multimodal Learning

This type of learning consists in teaching a model to recognise features from text, images, or audio. 

Gato, proof of the capabilities of machine learning

Gato is an algorithm developed by DeepMind that learns how to play a wide variety of games. Using a neural network and tree search algorithm, Gato is able to make decisions and improve its play style.

Gato was quickly able to outperform the best video game players. This algorithm points to the strength of machine learning to interpret images and sound and learn from them. It attests to machine learning’s capability to solve complex problems.

Artificial intelligence has come a long way since the 1940s. Increasingly complex data models have enabled artificial intelligence to do more than just react to data. Indeed, models have been developed that can make decisions or predictions based on inputted data. Artificial intelligence is now capable of solving very complex problems at astronomical speed.