Back-to-Basics: Explaining Artificial Intelligence

You would have to make an effort not to hear about the hype surrounding AI and its many uses, both real and potential.

Some of the most hyped uses of AI, like ChatGPT, seems to have sprung up from nowhere, but in reality it is the culmination of many years of development and thousands of hours of running expensive equipment to process data. Understanding how it works could go a long way to calming some of the fears surrounding AI, that it will do everything from take our jobs to take our lives.

Underneath it all, AI systems have many moving parts that have to work together. It was the cumulative breakthroughs of multiple parts that led to the current AI summer as the technological capabilities catch up with the field's ambitions. So let’s break it down.

What is AI?

Artificial intelligence is the ability of a computer or computer-controlled device to perform tasks commonly associated with intelligent beings, such as the ability to learn and perform complex tasks.

Despite the fears that it will grow a mind of its own, an artificial intelligence system has no initiative. It does what it is programmed to do and when it is told. A different technology called artificial general intelligence (AGI) has more initiative and acts on what it learns. But AGI is many years away even by the most optimistic of projections.

There are many different types of AI, but they all share a common goal of automating a system through learning and eliminating the potential for error. Some of the most common AI techniques include the following:

Machine learning: Machine learning (ML) lets systems learn from data. Machine learning algorithms are trained on a large amount of data, and they can then use this data to recognize everything from a potential hacker to image recognition to making predictions or decisions.

Natural language processing: Natural language processing (NLP) lets systems understand and process human language input. Input, such as searching and a search engine, is usually done through a series of keywords. NLP systems operate on full sentence structures to derive context.

Computer vision: Computer vision lets systems see and recognize images and objects from the world around them. It is used in everything from security systems for intrusion detection to self-driving cars.

Models make the world go around

AI operates on what are called models. An AI model is a computer program that is trained to perform a specific task, such as image recognition, natural language processing, fraud detection or others.

AI models are trained on large amounts of data. The data is used to teach the model how to perform its designated task. Once an AI model has been trained, it can be used to make predictions on new data.

They are taught in one of three ways: supervised learning, where models are trained on labeled data, or data that has been tagged with the correct answer; unsupervised learning models, which are trained on data that has not been tagged with the correct answer; and reinforcement learning, where models are trained by trial and error and a reward is given for taking actions that lead to desired outcomes.

The machine learning two-step

Machine learning is a two-part process by which models are made, consisting of training and inference. Training is the process of teaching a machine learning model to perform a task, such as image recognition, text recognition or spotting behavioral abnormalities. This is done by feeding the model a large amount of data and allowing it to learn from the data. The larger the data set, the more accurate the match – in theory, anyway.

This is by far the most process-intensive portion, where the fastest, most powerful processors are used for the training portion. Depending on the size of the training parameters, training can take weeks or even months of running high performance processors at maximum speed and utilization.

Inference is the process of using a trained machine learning model to make predictions on new data. This is done by feeding the model the new data and asking it to make a prediction. The model will then use its knowledge of the data it was trained on to make a prediction about the new data.

The inference part is much easier. Inference is matching input with the training. For example, in facial recognition, the model will be trained on thousands if not millions of faces, and inference matches a new face to those in the existing facial database.

Generative AI is one of the most discussed types of AI at present. Sometimes it is put in the same category as conversational AI, because it is used for human-machine interaction.

Older chatbots operate by recognizing keywords in user input and matching them to pre-written or existing response texts. With generative AI, the answer is generated on the spot by the AI system. You can ask the same question over and over again and get a slightly different answer every time.

The Importance of the GPU

Of all the hardware advances, the advent of the GPU as an AI processor has been the most significant. GPUs have been around for decades, but were dedicated to accelerating 3D graphics in videogames.

The key feature of GPUs is that they are massively parallel floating-point processors. Floating-point math is how 3D graphics are rendered in size, color, light, and shape. It also happens that floating-point math is vital to AI because it is the most efficient way to represent numbers in computers.

Floating-point numbers are more efficient than integers because they can represent a wider range of values. Integers can only represent whole numbers, while floating-point numbers can represent both whole numbers and fractionals.

About 20 years ago, Nvidia recognized this potential and began positioning its GPU technology not only as a means for acceleration of 3D games, but for AI processing. It developed a programming language called CUDA that allowed developers to program specifically to its GPU, and has made a 20-year effort to get CUDA taught in computer science schools around the world.

This is why Nvidia is so strongly linked to AI. Its only real competitor in the GPU space, AMD, was struggling just to survive for many years. The company has turned the corner and is now making a concerted effort to be more of a participant in the AI race, but it has a long way to catch up.

Algorithms are everything

There is a wide range of programming languages used in AI development: Python, C++, Java, R, and LISP, among others. Programmers generally work with the language that they know best, and Python is the most popular because it is the easiest to learn and there is a very large library of existing code for developers to use.

What’s far more important are the algorithms used in AI. An algorithm can be defined as a set of rules to be followed when performing calculations or some other task-oriented operations, especially by a computer. These existed long before AI came to the fore. For example, every time you make a purchase on Amazon, it makes recommendations for other products that may interest you. That’s the result of an algorithm.

Algorithms fall into several categories:

Problem-solving: These algorithms provide a systematic step-by-step procedure and logical rules necessary to achieve a desired outcome. These are used in tasks like pattern recognition, data analysis, decision-making, and predictions.

Learning and adaptation: These algorithms allow AI models to recognize patterns, make predictions, and adapt to changing conditions without being explicitly programmed for every possible scenario.

Decision-making and optimization: These algorithms are used to make intelligent decisions and optimize processes in AI systems based on the available data. An example would be the best route displayed on a GPS system.