What is AI?
Artificial intelligence (AI) lets computers and machines to resemble human learning, comprehension, problem solving, decision-making, creativity, and autonomy.
AI-enabled devices and apps can observe and recognise items. Human language can be understood and responded to by them. They are able to pick up new skills and knowledge. They are able to offer consumers and specialists thorough recommendations. They may behave on their own without human intelligence or involvement (a self-driving car is a classic example).
In 2024, most AI researchers and practitioners and most AI-related news storiesfocus on generative AI, or “gen AI,” a system that can generate original text, images, movies, and other content. One must understand machine learning (ML) and deep learning, the technology behind generative AI tools, to understand it.
Artificial Intelligence models
Machine learning
Artificial Intelligence (AI) can be simply understood as a collection of nested or derivative notions that have developed over more than 70 years:
Machine learning, which lies directly beneath artificial intelligence, is the process of building models by teaching an algorithm to make judgements or predictions using data. It includes a wide variety of methods that let computers learn from and draw conclusions from data without needing to be specifically programmed for a given activity.
There are numerous kinds of machine learning algorithms or approaches, such as clustering, k-nearest neighbour (KNN), logistic regression, decision trees, random forests, support vector machines (SVMs), and more. These methods can be applied to various types of data and challenges.
Neural networks, often known as artificial neural networks, are among the most widely used classifications of machine learning algorithms. The structure and operations of the human brain are modelled by neural networks. Neural networks process and analyse complicated data by utilising interconnected layers of nodes, which are similar to neurones. Neural networks are ideal for jobs involving the discovery of intricate linkages and patterns in vast volumes of data.
Supervised learning is the most basic type of machine learning, where algorithms are trained using labelled data sets to correctly classify data or predict outcomes. Each training example is paired with an output label by humans in supervised learning. The model is intended to be able to predict the labels of fresh, unseen data by learning the mapping between inputs and outputs in the training data.
Deep learning
Multilayered neural networks, or deep neural networks, are used in deep learning, a branch of machine learning, to more accurately mimic the intricate decision-making processes of the human brain.
Unlike neural networks used in traditional machine learning models, which often contain only one or two hidden layers, deep neural networks have an input layer, at least three, but typically hundreds of hidden levels, and an output layer.
Unsupervised learning is made possible by this many layers, which can automatically extract features from sizable, unlabelled, and unstructured data sets and guess what the data might represent.
Large-scale machine learning is made possible by deep learning since it eliminates the need for human interaction. Its various applications include natural language processing (NLP), computer vision, and other jobs that need fast and accurate pattern and correlation recognition in large data sets. Deep learning powers most AI applications it use daily.
Deep learning allows:
- Using both labelled and unlabelled data, semi-supervised learning blends supervised and unsupervised learning to train AI models for tasks like regression and classification.
- Instead than depending on labelled data sets to provide supervisory signals, self-supervised learning leverages unstructured data to produce implicit labels.
- Reinforcement learning is learning by using reward systems and trial-and-error instead than deriving knowledge from hidden patterns.
- In transfer learning, model performance is enhanced on a related task or distinct data set by applying knowledge acquired from one task or data set.
Generative AI
Sometimes referred to as “gen AI,” generative AI refers to deep learning models that can produce complex original material in response to a user’s prompt or request. Examples of this type of content include realistic video or audio, long-form text, and high-quality photographs.
In essence, generative models use a reduced representation of their training data to generate new work that is comparable to but distinct from the original data.
Statistics has long employed generative models to examine numerical data. The past ten years, however, have seen them develop to analyse and produce increasingly complex data kinds. At the same time as this development, three advanced deep learning model types emerged:
- Models were able to produce several versions of material in response to a prompt or instruction because to the introduction of variational autoencoders, or VAEs, in 2013.
- In order to create creative images in response to prompts, diffusion models which were originally introduced in 2014 add “noise” to photographs until they are unrecognisable and then remove the noise.
- Transformers, sometimes known as transformer models, are trained on sequenced data to produce lengthy content sequences (e.g., commands in software code, forms in an image, words in phrases, or frames in a video). Most of today’s headline-grabbing generative AI technologies, such as ChatGPT and GPT-4, Copilot, BERT, Bard, and Midjourney, are built using transformers.
How generative AI works
Generative AI generally functions in three stages:
- Training in order to establish a foundational model.
- Tuning, which modifies the model to fit a particular use case.
- Generation, assessment, and additional fine-tuning to increase precision.
Training
A deep learning model known as the “foundation model” is the starting point for generative AI and forms the basis of several generative AI application types.
Large language models (LLMs), developed for text generating applications, are currently the most often used foundation models. However, there are also multimodal foundation models that support many types of content, as well as foundation models for the generation of images, videos, sounds, or music.
Practitioners use vast amounts of pertinent raw, unstructured, unlabelled data terabytes or petabytes of text, photos, or video from the internet to train a deep learning algorithm in order to build a foundation model. A neural network with billions of parameters encoded representations of the entities, patterns, and relationships in the data is produced by the training process and is capable of producing material on its own when given instructions. The fundamental model is this one.
This training procedure is costly, time-consuming, and computationally demanding. It takes weeks to process and thousands of clustered graphics processing units (GPUs), which can cost millions of dollars. This step and its expenses can be avoided by gen AI developers with open source foundation model initiatives like Meta’s Llama-2.
Tuning
The model then needs to be adjusted for a particular content creation requirement. There are several ways to accomplish this, including:
- Feeding the model application-specific labelled data queries or prompts that the application is expected to receive, along with the matching right responses in the desired format, is known as fine-tuning.
- By using human users to assess the precision or applicability of model outputs, reinforcement learning with human feedback (RLHF) allows the model to self-improve. This can be as easy as having users type or speak corrections back to a virtual assistant or chatbot.
Creation, assessment, and additional fine-tuning
Users and developers evaluate the results of their generative AI applications on a regular basis, and they can adjust the model as frequently as once a week to improve accuracy or relevance. The foundation model itself, on the other hand, is revised considerably less frequently possibly once every 18 months or once a year.
Retrieval augmented generation (RAG), a method of expanding the foundation model to leverage pertinent sources outside of the training data to adjust the parameters for increased accuracy or relevance, is an additional choice for enhancing the performance of a gen AI software.
