Artificial intelligence: what it is and how it works

What is artificial intelligence?

Artificial intelligence (AI) is a broad field of computer science focused on creating smart machines capable of performing intellectual tasks.

There are many approaches to building such algorithms. Advances in machine learning and deep learning in recent years have significantly reshaped the technology industry.

What definitions of artificial intelligence exist?

The English mathematician Alan Turing set out AI’s fundamental goal and vision in his 1950 paper “Computing Machinery and Intelligence”. He posed a simple question: “Can machines think?” He also proposed the famous test that bears his name.

At its core, AI is a branch of computer science that seeks to answer Turing’s question in the affirmative—an attempt to reproduce or simulate human intelligence in machines.

The ultimate aim of AI still prompts questions and debate. Defining AI simply as “intelligent machines” is limiting, not least because neither scientists nor philosophers can pin down what intelligence is, or what exactly makes a machine smart.

Stuart Russell and Peter Norvig, authors of the textbook “Artificial Intelligence: A Modern Approach”, organise their work around intelligent agents in machines and define AI as “the study of agents that receive percepts from the environment and take actions”.

In a 2017 talk at Japan AI Experience, DataRobot chief executive Jeremy Achin opened with the following description of how AI is used today:

“AI is a computer system able to perform tasks that require human intelligence … Many of these systems run on machine learning, others on deep learning, and some of them on very boring things such as rules”.

While these definitions may seem abstract, they help delineate core lines of theoretical research in computer science and offer concrete pathways for deploying AI programmes to solve practical problems.

What is Turing’s contribution to AI?

In the mid-20th century Alan Turing laid the theoretical foundations that were ahead of their time and underpinned modern computer science, earning him the moniker “the father of computer science”.

In 1936 he devised an abstract computing device—the Turing machine—an important component of algorithmic theory that underlies modern computers. In theory, such a machine can solve any algorithmic problem.

In turn, if an algorithm can run on a Turing machine, the programming language used to create it is said to be “Turing-complete”, meaning any algorithm can be written in it. For example, C# is Turing-complete; HTML is not.

His name is also attached to a thought experiment unrelated to the machine but directly tied to AI—the Turing test. In academic circles, once a machine passes this test, one could meaningfully talk about the emergence of intelligent machines.

The idea is simple: a person interacts via text with both a machine and a human. The computer’s task is to mislead the participant and convincingly pass for a person.

What kinds of AI are there?

AI is commonly divided into two broad categories:

• Weak AI [Narrow AI]: operates in a limited context and simulates human intelligence. It is often designed to perform a single task very well. Although such systems may appear smart, they work under tight constraints.
• Artificial general intelligence (AGI): sometimes called “strong AI”, the kind seen in films—robots from “Westworld” or the hologram Joi from “Blade Runner 2049”. AGI is a machine with general intelligence that, like a human, can apply it to any problem.

What is weak AI?

Weak AI is everywhere and, to date, is the most successful form of artificial intelligence. 

Focused on specific tasks, it has delivered numerous breakthroughs over the past decade that have “produced significant societal benefits and contributed to the nation’s economic vitality”, according to the 2016 Obama administration report “Preparing for the Future of Artificial Intelligence”.

Examples of weak AI include:

• Google search;
• image-recognition software;
• Siri, Alexa and other voice assistants;
• self-driving cars;
• recommendation systems at Netflix and Spotify;
• IBM Watson.

How does weak AI work?

Most weak AI builds on advances in machine learning and deep learning. The terms can be confusing, but they are distinct. Venture capitalist Frank Chen proposed the following:

“Artificial intelligence is a set of algorithms that try to imitate human intelligence. Machine learning is one of them, and deep learning is one of the methods of machine learning”.

In other words, machine learning feeds data to a computer and uses statistical methods to help it learn to perform tasks without being explicitly programmed, avoiding the need for millions of lines of hand-written code. Popular paradigms include supervised learning (with labelled datasets), unsupervised learning (with unlabelled datasets) and reinforcement learning.

Deep learning is a type of machine learning in which inputs are processed through a neural-network architecture inspired by biology. 

Neural networks contain a series of hidden layers through which data are processed, allowing the machine to “go deeper” in its learning, form connections and weight inputs to achieve better results.

What is machine learning?

Artificial intelligence and machine learning are not the same. Machine learning is a subfield of AI.

The most common types are supervised, unsupervised and reinforcement learning.

Supervised learning

Used when developers have a labelled dataset and know which features an algorithm should look for.

It typically falls into two categories: classification and regression.

Classification applies when objects must be assigned to predefined classes. It is used in spam filters, language detection and flagging suspicious transactions.

Regression is used when an object must be mapped onto a timeline—for example, forecasting the price of securities, demand for a product or making medical diagnoses. 

Unsupervised learning

A less popular form owing to its unpredictability. Algorithms train on unlabelled data and must find features and patterns on their own. It is often used for clustering, dimensionality reduction and association discovery.

Clustering is like classification but without known classes. The algorithm must identify similarities among objects and group them into clusters. It is used to analyse and label new data, compress images or merge map labels.

Dimensionality reduction generalises specific features into higher-level abstractions. It is often used to determine the topics of texts or build recommendation systems.

Associations are used in marketing—for example, when designing promotions and sales or analysing user behaviour on a website. They can also underpin a recommender system.

Reinforcement learning

This trains an agent to survive in its environment—anything from a video game to the real world. 

For instance, there are algorithms that play Super Mario as well as humans; in the real world, Tesla’s autopilot or a robot vacuum cleaner does its best to avoid obstacles.

Reinforcement learning rewards an agent for correct actions and penalises mistakes. The algorithm need not memorise its entire experience or enumerate all possible futures; it must learn to act according to the situation. 

Remember when a machine beat the world champion at Go? Long before that, scientists established that the number of possible moves in Go exceeds the number of atoms in the universe. No existing program could calculate every possible line. Yet AlphaGo, Google’s algorithm, managed the feat without searching all moves in advance, acting on the circumstances—and with remarkably high accuracy.

What are neural networks and deep learning?

The idea of artificial neural networks is not new. It was first formulated by American researchers Warren McCulloch and Walter Pitts in 1943.

Any neural network consists of neurons and connections between them. A neuron is a function with many inputs and one output. Neurons exchange information through channels, each with a certain weight.

A weight is a parameter that determines the strength of a connection between neurons. A neuron does not interpret the content it sends; the weight governs which inputs to respond to and which to ignore.

For example, if a neuron outputs the number 50 and the connection’s weight is set to 0.1, the result will be 5.

As architectures grew more complex, neurons were connected in layers. Within a layer, neurons do not interact with each other; they receive information from the previous layer and pass it to the next.

As a rule, the more layers a network has, the more complex and accurate the model. Fifty years ago researchers hit the limits of computing power. The technology disappointed and was shelved for many years.

It returned to prominence in 2012, when University of Toronto students Alex Krizhevsky, Ilya Sutskever and Geoff Hinton won the ImageNet computer-vision competition. They used a convolutional neural network for image classification, achieving an error rate of 15.3%—more than ten percentage points lower than the second-place team. The revolution in deep learning was powered in large part by advances in graphics processors.

Deep learning differs from generic neural networks primarily in methods for training very large networks. In practice, developers rarely draw a hard line between “deep” and other networks; even networks with five layers are built with deep-learning libraries such as Keras, TensorFlow or PyTorch.

Today the most popular architectures are convolutional neural networks (CNNs) and recurrent neural networks (RNNs).

CNNs are often used for face recognition, object detection in photos and video, image enhancement and related tasks. RNNs are used in machine translation and speech synthesis. Since 2016 Google Translate has run on an RNN architecture.

Generative adversarial networks (GANs) have also become popular. They pit two neural networks against each other: one generates data—for example, an image—while the other tries to distinguish real samples from fakes. Because the two networks compete, this forms an antagonistic game between them.

GANs are widely used to create photorealistic images. For example, the repository This Person Does Not Exist consists of portrait photos of “people” created by a generative network.

What is artificial general intelligence?

Building a machine with human-level intelligence that can be applied to any task is the Holy Grail for many AI researchers, but the hunt for AGI is fraught with challenges.

AGI has long inspired dystopian science fiction in which superintelligent robots overwhelm humanity, but experts broadly agree this is not something to fret about in the near term.

American inventor and futurist Ray Kurzweil predicted AGI would arrive by 2029. His colleague Rodney Brooks is less sanguine, arguing that a tipping point in machine intelligence will come by 2300.

Stuart Russell, a co-author of “Artificial Intelligence: A Modern Approach”, suggests AGI may emerge by accident, as nuclear energy did in 1933. For such a volatile, poorly understood technology, he argues, making forecasts is largely futile.

Subscribe to ForkLog on Telegram: ForkLog AI — all the news from the world of AI!