Want to know AI terminology? From NLP to neural networks, learn the basics

Algorithm: Today an algorithm is usually a set of instructions for a computer to follow. Those designed to search and sort data are examples of computer algorithms that work to take information and arrange it in a particular order. It can consist of words, numbers, codes and symbols, as long as they detail finite steps to complete a task. But algorithms have ancient origins, going back at least to the Babylonian clay tablets. The Euclidean algorithm of division is still in use today, and even brushing your teeth can be distilled into an algorithm, although it is quite complex given the fine-grained coordination of movements that goes into daily rituals.

Machine learning: A field of AI that relies on techniques that enable computers to learn from the data they process. Scientists have previously attempted to create artificial intelligence by programming knowledge directly into computers.

You can feed your ML system millions of animal photos from the web and label each one as a cat or dog. The process of imparting this information is known as “training.” Even without knowing anything else about the animal, the system can identify statistical patterns in photos and use those patterns to recognize and classify new examples of cats and dogs.

ML systems are very good at recognizing patterns in data, but they are less effective when the task requires long chains of inference or complex planning.

Natural language processing: A form of machine learning that can interpret and respond to human language. It powers Apple's Siri and Amazon.com's Alexa. Many of today's NLP techniques select sequences of words based on their probability of achieving a goal, such as summarization, question-and-answer, or translation, said Daniel Mankowitz, a researcher at DeepMind, a Google subsidiary that researches artificial intelligence. It is said that they are doing so. .

The context of the surrounding text tells you whether the word “club” refers to sandwiches, golf, or nightlife. The roots of this field go back to his 1950s and his 1960s. At that time, processes were taking place to help computers analyze and understand. To use the language, scientists themselves had to code the rules. Today, computers are trained to make these language associations on their own.

Neural network: A machine learning technique that mimics the way neurons operate in the human brain. In the brain, neurons can send and receive signals that drive thoughts and emotions. In artificial intelligence, groups of artificial neurons or nodes similarly send and receive information from each other. Artificial neurons are essentially lines of code that act as connection points with other artificial neurons to form a neural network.

Unlike older forms of machine learning, it always trains on new data and learns from its mistakes. For example, Pinterest uses neural networks to process large amounts of data about its users, including their searches, the boards they follow, and the Pins they click and save to create images and images that catch consumers' attention. Find ads. At the same time, the network examines your advertising data (such as what kind of content makes you click on ads) to learn your interests and serve you more relevant content.

Deep learning: A type of AI that employs neural networks to continuously learn. In deep learning, “deep” refers to multiple layers of artificial neurons in a network. Compared to neural networks, which are better at solving smaller problems, deep learning algorithms are capable of more complex processing because of their interconnected layers of nodes. In a 2019 paper, Oxford University doctoral candidate David Watson wrote that they were inspired by the structure of the human brain, and that neural networks are weaker when compared to the performance of real human brains. Since then, the method has exploded in popularity, calling it inefficient and short-sighted. This is his groundbreaking paper published in 2012 by three researchers at the University of Toronto.

Large-scale language models: Deep learning algorithms that are trained on vast amounts of data so they can summarize, create, predict, translate, and synthesize text and other content. A common starting point for programmers and data scientists is to train these models on publicly available open-source data sets from the internet.

LLM is derived from the “Transformer” model developed by Google in 2017, which makes it cheaper and more efficient to train models using vast amounts of data. . His first GPT model for OpenAI, released in 2018, was built on the work of Google's Transformers. (GPT stands for Transformer) Known as a multimodal language model, LLM can operate with different modalities such as language, images, and audio.

Generative AI: A type of artificial intelligence that can create different types of content such as text, images, video, and audio. Generative AI is the result of humans inputting information or instructions called prompts into a so-called underlying model, which then generates an output based on the given prompts. Foundation models are a class of models trained on vast and diverse amounts of data that can be used to develop more specialized applications, such as chatbots, coding assistants, and design tools. Such models and their applications include text generators such as OpenAI's ChatGPT and Google Bard, as well as image generators such as OpenAI's Dall-E and Stability.ai's Stable Diffusion.

The release of ChatGPT last November sparked an explosion of interest in generative artificial intelligence. This makes it easy to interact with OpenAI's underlying technology by typing questions and prompts in everyday language. Similarly, OpenAI's Dall-E 2 creates realistic-looking images.

Such models are trained not only on the Internet but also on more customized datasets to find long-range patterns in a set of data, allowing the AI ​​software to choose the appropriate next word to write or create. You will be able to express sentences and paragraphs.

Chatbot: A computer program that can talk to people in human language. Modern chatbots rely on generative AI, allowing people to ask questions and provide instructions to the underlying model in human language. ChatGPT is an example of a chatbot that uses a large language model (in this case his OpenAI's GPT). People can converse with ChatGPT on topics ranging from history to philosophy, ask it to generate lyrics in the style of Taylor Swift or Billy Joel, or suggest edits to computer programming code. . ChatGPT can synthesize and summarize vast amounts of text and convert it into human language output on any topic that currently exists in the language.

Hallucination: Presented as such when the underlying model produces a response that is not based on facts or reality. Hallucinations are different from bias, which is another problem that occurs when the training data has biases that affect the output of her LLM. Hallucinations are one of the major drawbacks of generative AI, and many experts call for human supervision of her LLM and its output.

The term was popularized in a 2015 blog post by OpenAI founding member Andrej Karpathy. He writes about how models can “illusion” text responses, such as fabricating plausible mathematical proofs.

Artificial General Intelligence: A virtual form of artificial intelligence that allows machines to learn and think like humans. While the AI ​​community has not reached a broad consensus on what AGI entails, Ritu Jyoti, a technology analyst at research firm IDC, believes that AGI helps solve problems, adapt to the environment, and He said self-awareness and consciousness would be required to be able to perform the functions. Scope of the task.

Companies including Google DeepMind are working on developing some form of AGI. DeepMind said the AlphaGo program aired numerous amateur games, which helped deepen human understanding of rational games. Since then, I have played against different versions of myself thousands of times, learning from my mistakes each time.

Over time, AlphaGo improved and became better and better at learning and making decisions (a process known as reinforcement learning). DeepMind said its MuZero program subsequently mastered Go, Chess, Shogi, and Atari without needing to be taught the rules, demonstrating its ability to plan winning strategies in an unknown environment. Some see this progress as an incremental step in the direction of AGI.

Email Steven Rosenbush (steven.rosenbush@wsj.com), Isabelle Bousquette (isabelle.bousquette@wsj.com), and Belle Lin (belle.lin@wsj.com).