Fundamentals of AI
So how exactly does AI work?
Firstly, an AI system receives input data in various forms such as speech, text, images and so on. It then proceeds to analyse this data by employing a set of rules and algorithms. These rules and algorithms, interpret, predict, and take action based on the input information. After the processing phase, the system produces an outcome, which can be classified as either a success or a failure with reference to the input data. The outcome produced by the system gets evaluated through certain mechanisms of analysis, discovery, and feedback. Finally, the system utilizes these evaluations to make adjustments to the input data, rules, algorithms, and target outcomes. This process continues as a loop until the desired result is achieved.
In order for the system to be classified as AI (Artificial Intelligence), all of its constituent elements must operate in harmony.
Before we start discussing the core concepts, let’s take a look at a couple of key terminologies that are recurrent in the field of AI.
The first one is algorithms.
Algorithms
In the field of mathematics and computer science, an algorithm is a finite series of precise instructions, typically designed to solve a particular set of problems or to carry out a computational task. Algorithms serve as explicit guidelines for executing calculations and processing data.
In basic terms, algorithms are like step-by-step instructions that guide a computer through calculations. You can think of them as instruction manuals that tell the machine precisely what to do and when. Without this guidance, the computer might struggle to begin or proceed with tasks. Even in systems designed for machine learning, algorithms provide the initial direction, serving as a starting point for the programming process. In essence, they are the foundation for computer operations.
Simple algorithms consist of a series of steps designed to carry out a straightforward and predictable task. While these algorithms may seem sophisticated, they do not fall under the category of artificial intelligence (AI). For instance, Facebook employs such algorithms to target individuals with advertisements based on their online browsing and shopping history.
On the other hand, complex algorithms, often discussed in the context of AI, encompass areas such as machine learning, deep learning, neural networks, and fuzzy logic. In this context, machine learning refers to computer systems capable of learning from past inputs and outcomes to optimize future decisions, as seen in applications like facial recognition.
However, it's important to note that the term "complex algorithms" is sometimes used interchangeably with AI, but AI is a broader and more intricate field than what these algorithms alone represent.
The next key terminology that we are at today is data.
Data
Data is the fuel that powers artificial intelligence. Building an AI system typically involves sourcing substantial volumes of data and generating datasets for various purposes, including training, testing, and evaluation, as well as eventual deployment. This process is repetitive in the sense that it may require several rounds of training, testing and evaluation until the desired outcome is achieved and data plays an important role at each step.
To embark on AI projects, one must first ensure the right data is available in sufficient quantities. This encompasses data acquisition, cleansing, and organization, all crucial steps within the data journey. Key considerations include proper sampling techniques to represent the real world, understanding how to obtain diverse data types, maintaining data completeness, ensuring consistency, and enriching data quality.
Moreover, addressing bias in data is important to prevent AI from perpetuating inequities. Data is not just a precursor but a fundamental component throughout the AI system’s development cycle, and careful attention to data quality and bias mitigation is essential for successful AI implementation.
With that understanding, Let's move on to discuss some essential components of AI for a better understanding of how these systems are built and trained.
The five key components of AI which are crucial to understand how it works is as follows:
Machine Learning
Deep Learning
Neural Network
Natural Language Processing
Computer vision
Cognitive computing
Machine learning is a central principle to the concept of AI. Machine learning is crucial to comprehend how AI models’ function and to understand the mechanisms behind the creation of AI tools. AI systems are built on algorithms that enable computers to learn from data and improve their performance over time. This process is referred to as machine learning.
Machine Learning, as a subset of AI, uses computer algorithms to analyze data and make intelligent decisions based on what it has learned. Instead of following rules-based algorithms, machine learning builds models to classify and make predictions from data. Machine Learning takes data and answers and creates the algorithm. Instead of getting answers in the end, we already have the answers to problems. What we get is a set of rules that determine what the machine-learning model will be. The model determines the rules, and the if-then-else statement when it gets the inputs. This model, unlike a traditional algorithm, can be continuously trained and used in the future to predict values or answers. Machine Learning relies on defining behavioural rules by examining and comparing large datasets to find common patterns.
To ensure the effectiveness of machine learning, it needs an appropriate training dataset or large collection of data. A dataset is a collection of various types of data stored in digital format, serving as a fundamental component of any machine learning project. Datasets primarily consist of a range of data formats including images, texts, audio, videos, numerical data points, and so on. These datasets are crucial for addressing various Artificial Intelligence challenges such as image or video classification.
This dataset is used to train machine learning algorithms by providing examples from which the model learns patterns and relationships, such as, for example, identifying what constitutes the appearance of an animal such as a cat, or the classification of objects into specific groups and so on.
Following this training, the system generates a mathematical model that represents the relationships and patterns it studied from the dataset during the learning phase. This mathematical model is then used to make predictions or decisions based on the acquired knowledge.
AI systems that we encounter these days such as OpenAI’s GPT-4 or Google Bard have specific AI models underlying their systems. Essentially, they function as high-powered prediction machines, similar to how autocomplete works on your smartphone. After selecting the relevant data for training the model, researchers undertake a pre-processing phase, which involves cleaning, transforming, and normalizing the data to prepare it for use in machine learning algorithms. Additionally, they address issues such as missing values, outliers, and inconsistent formatting. It's important to note that the quality and quantity of data fed into the system, significantly impact the performance of these models. Any biases, limitations, gaps, or flaws present in the data are likely to manifest in the trained models as well. And considering how the data may be deficient is a critical aspect of evaluating different models thoughtfully.
Once these models are developed, they are deployed into applications or systems to make predictions or decisions based on new, previously unseen data. This is the point at which most users, like us, who may not have a background in computer science, engage with AI systems.
The stage where models are built from machines that use algorithms to learn from data can involve various training methods. Some of these methods are as follows:
Supervised learning
Unsupervised learning
Reinforcement learning
Deep learning
Transfer learning
1. Supervised learning: In supervised learning, algorithms acquire knowledge from human-labelled data in order to make predictions or classifications. For instance, a model could be exposed to thousands of images that are annotated with the names of the animals they depict, allowing the model to learn how to classify images based on the animals they contain that it might encounter in the future.
For instance, we can provide a machine learning program with a vast collection of bird pictures and train it to label any picture of a bird as "bird." Similarly, we can create a label for "cat" and provide cat pictures for training. When the machine model encounters a picture of a cat or a bird, it labels it with a certain level of confidence. The more samples you provide to a supervised learning algorithm, the more accurate it becomes at classifying new data. This approach is effective but is dependent on the availability of accurately labelled data.
2. Unsupervised learning:
In contrast to supervised learning, unsupervised learning involves the use of algorithms that identify patterns or structures within unlabelled data. This approach is often employed when there is no clear understanding of what to expect from the data or when the data is so intricate that manually labelling it would be impractical or even impossible. Unsupervised learning algorithms are designed to autonomously detect concealed patterns, clusters, or associations within the data.
For instance, imagine providing an algorithm with a collection of various books, with the task of grouping them based on their subjects, reading levels, and languages. A clustering algorithm could analyse all the books, decode which ones share similarities in terms of these attributes, and then classify them accordingly.
Unsupervised Learning involves giving the algorithm unlabelled data and letting it discover patterns independently. In this case, you provide input but not labels, allowing the machine to identify qualities and draw connections based on the unlabelled data.
3. Reinforcement learning: Reinforcement learning is a specific branch of machine learning in which a system learns to adopt particular behaviours through a process of taking actions and observing the outcomes or rewards associated with those actions. The objective is for the system to learn which actions produce the highest rewards. For instance, reinforcement learning has found applications in gaming, where AI systems learn to excel at a game by playing it repeatedly and refining their strategies based on whether they win or lose previous matches. Additionally, reinforcement learning can involve human reinforcement learning, in which the AI generates output, and a human evaluates the quality of that output to help in training the model. Many of the most widely used generative AI models incorporate some form of human reinforcement learning in the process of their development.
Reinforcement Learning, operates by giving an algorithm a set of rules, constraints, and goals, allowing it to learn how to achieve those goals. The algorithm determines the best course of action within the given constraints, aiming to maximize success and reduce failure. This approach is suitable for teaching machines to perform tasks such as playing chess or navigating obstacle courses.
4. Deep learning:
Deep learning is a subset of machine learning that employs layered algorithms to create Neural Networks, which are artificial systems designed to replicate the structure and functioning of the human brain. This approach allows AI systems to continually learn and enhance the quality and accuracy of their outcomes. It is this capability that empowers these systems to learn from unstructured data sources like photos, videos, and audio files.
For instance, Deep Learning plays a pivotal role in enabling natural language understanding in AI systems, helping them identify context and intent in communication. Unlike traditional machine learning, deep learning algorithms do not directly map input to output. Instead, they rely on multiple layers of processing units. Each layer processes the output of the preceding one, forming a chain of transformations. This multi-layered structure is why it's called "deep" learning.
When crafting deep learning algorithms, developers and engineers configure the number of layers and the types of functions that connect the outputs of each layer to the inputs of the next one. They then train the model by providing it with numerous annotated examples. For instance, a deep learning algorithm could be exposed to thousands of images along with labels corresponding to each image's content. The algorithm processes these examples through its layered neural network and adjusts the weights of the variables in each layer to recognize common patterns that define images with similar labels.
One of the key advantages of Deep Learning is its ability to tackle a major issue seen in older generations of learning algorithms. While the performance of traditional machine learning algorithms tends to slow down as datasets grow larger, deep learning algorithms continue to improve with more data. Deep Learning has proven highly effective in a range of tasks, including image captioning, voice recognition and transcription, facial recognition, medical imaging, and language translation. Additionally, it is a critical component of driverless cars and autonomous vehicle technology.
Neural networks
Neural networks are technologies designed to mimic the structural principles of natural brains. They embody a bottom-up approach to artificial intelligence, underlined by the belief that intelligence emerges from the brain's architecture.
Neural networks learn through a process known as backpropagation, which relies on a set of training data that maps known inputs to desired outputs. The process begins with input data being fed into the network, which then produces outputs. An error function calculates the difference between the given output and the desired output. Adjustments are made to minimize these errors.
Picture these networks as a system of interconnected stations, much like a well-organized train network. There's an initial station, called the input layer, numerous intermediate stations known as hidden layers, and a terminal station called the output layer. Each station, or "neuron," holds and processes information. Data flows from the starting station, through the intermediate stations, and ultimately arrives at the final station.
Along this journey, the information undergoes transformations and processing based on rules the system learns from the provided data. While neural networks aren't a new concept, their recent accomplishments are largely attributed to the accessibility of vast datasets, innovative algorithms, and enhanced computational capabilities.
Neural networks have emerged as a key driving force behind the recent achievements in machine learning, which, in turn, is the primary catalyst for the ongoing wave of progress in AI and serve as the fundamental building blocks of deep learning for tackling complex tasks.
Deep learning’s reliance on various types of neural network has made it a technique of significant importance in the field of artificial intelligence. It has played a pivotal role in recent achievements across a wide range of applications. These include advances in speech recognition, visual object recognition, object detection, and extends its influence into domains like drug discovery and genomics.
Deep learning models are engineered to AUTONOMOUSLY acquire data representations by training on large datasets. They excel at finding patterns in unstructured data types such as images, audio, and text. For instance, deep learning has been employed to automatically translate text between different languages.
Deep learning's ability to automatically learn complex patterns from data has made it a powerful tool with far-reaching implications in numerous fields.
5.Transfer learning: Transfer learning is a technique in which a pretrained model, initially trained for a specific purpose, is employed to handle a new but related problem. It involves taking an existing model, which has already been trained for one task, and reusing it as the starting point for a different, yet closely related, task. For example, one could train a neural network extensively on a large-scale image classification task and then fine-tune the same model to recognize specific types of objects in another task. The advantage of transfer learning is that it can significantly reduce computational time and enhance prediction performance, especially in scenarios where there is limited data, by exploiting the knowledge acquired during the initial task.
MACHINE LEARNING, despite its remarkable achievements, falls under the category of narrow AI. This approach excels at replicating highly SPECIFIC cognitive processes and tasks. It DOES NOT possess the inherent capability to easily transfer knowledge and insights from one domain to another. For instance, a machine learning system trained to differentiate between cats and dogs does not automatically possess the skill to understand natural language. While the underlying system may have the potential to be ADAPTED for various applications, it must undergo fresh training when used for new purposes. Each task typically requires specific training and tuning to achieve best results.
It's crucial to recognize which characteristics of machine learning have ethical implications.
The first feature we have to discuss is,
Opacity: Machine learning algorithms and neural networks are often so intricate that their internal workings are challenging to decipher, even for domain experts. These systems are dynamic and adaptive, continuously learning and evolving, making it difficult to achieve a comprehensive understanding of their inner mechanisms.
Unpredictability: moreover, due to this opacity, predicting the outputs of these systems based solely on an understanding of the inputs can be challenging, and in some cases, nearly impossible. The complex and evolving nature of machine learning models makes their behaviour difficult to anticipate their outcomes with certainty.
"Big data" requirements are indeed a significant characteristic of machine learning systems. They rely on extensive training datasets and substantial computational resources to build and train models effectively.
AI finds widespread applications in various domains, with natural language processing, speech, and computer vision being some of the most common areas of use. Let's look into each of these domains individually.
Humans possess the most advanced form of communication, which is natural language. While computers facilitate voice and text communication among humans, they do not inherently understand natural language.
Natural language processing (NLP), a subset of AI, enables computers to grasp the meaning embedded in human language. It focuses on the computational processing of natural languages. It seeks to replicate cognitive language abilities without true simulation of human cognitive processes. Instead, NLP employs a range of theoretically grounded computational techniques to analyse and represent naturally occurring texts at various levels of linguistic analysis. The ultimate goal is to achieve language processing capabilities similar to human comprehension for a multitude of problems and applications.
NLP harnesses machine learning and deep learning algorithms to decode the semantic meaning of words. This involves grammatical, relational, and structural analysis of sentences and an understanding of contextual usage of certain words with multiple meanings. NLP systems can also gauge intent and emotion, examining if a question arises from frustration, confusion, or irritation. By employing various linguistic models and algorithms, NLP systems make analyses about the user's language.
NLP relies on a combination of linguistic theories and technological tools, making it a dynamic field of both research and development. The texts analysed in NLP can encompass any human language, be it spoken or written, and can span diverse genres and modes of communication. The only requirement is that they be in a language used by humans to communicate to one another. Also, the text being analysed should not be specifically constructed for the purpose of analysis, but rather that the text is gathered from actual usage.
The concept of "levels of linguistic analysis" acknowledges that humans engage multiple types of language processing when using or understanding language. Humans employ various sentence structures to convey their thoughts, which are categorized by their construction and purpose. For instance, a simple sentence consists of a subject and predicate, while a compound sentence joins two or more independent clauses using conjunctions or semicolons. The purpose of a sentence determines its mood—declarative, imperative, interrogative, or exclamatory. Additionally, words within sentences serve distinct functions as parts of speech, with categories including nouns, pronouns, verbs, adjectives, adverbs, prepositions, conjunctions, and more in the English language. NLP aims to computationally dissect and comprehend these multifaceted aspects of human language.
The objective of natural language processing (NLP) is to develop a language comprehension and production framework to a level of precision where it becomes feasible for individuals to create computer programs capable of comprehending and generating natural language. At its core, NLP strives to achieve language processing that closely resembles human capabilities.
It's important to note that the term "processing" in NLP is chosen deliberately and should not be substituted with "understanding." This distinction arises because, despite its early designation as Natural Language Understanding (NLU), the field has not yet achieved the full goal of complete language understanding.
The backbone of modern NLP is, Large Language Models – these are deep learning models trained to produce text.
Large language models or LLMs are the underlying technology that powers text generators. An LLM is a centrally powered algorithm and composed of billions of parameters engaging with hundreds of billions of words. They have been integrated into many familiar platforms, such as translators, conversation bots, search, text suggestion, and as mentioned, text generators. For example, if you use Microsoft Office products, you may have noticed that Outlook and Word now provide suggested phrases to complete your text when you are writing, but if you are composing offline, this function disappears. Why does this occur? The suggested text is being created by powerful, distant computation using a Large Language Model. The most prevalent LLM being used currently is OpenAI’s GPT-3 and 4, and texts created by platforms using this LLM were making news for their fluency in recent years.
More recently, Meta (formerly Facebook) and Hugging Face (an LLM start up and OpenAI collaborator) have announced that they are releasing an equally powerful LLM -- but unlike GPT-3 which works with limited access for unvetted developers – they are publishing all of the accompanying documentation and parameters for public access; thus, anyone can now learn how to make their own LLM.
Another specialized field within artificial intelligence and computational linguistics is NATURAL LANGUAGE GENERATION (NLG). It focuses on developing computer systems capable of generating coherent and understandable text in human languages like English. NLG systems typically start with non-linguistic data as input and use their knowledge of language and the specific domain to automatically create various types of text, including documents, reports, explanations, and more.
NLG is not only a practical application but also a research area that delves into essential questions in AI, cognitive science, and human-computer interaction. Some of the fundamental issues it addresses include how to represent and reason with linguistic and domain knowledge effectively, what constitutes a well-written text, and the most effective ways to convey information between machines and humans.
Computer visions
Computer vision, a subset of artificial intelligence (AI), empowers computers and systems to extract meaningful insights from digital images, videos, and other visual inputs. It seeks to replicate aspects of the human visual system, enabling computers to recognize and process objects within images and videos, much like humans do. This technology facilitates the interaction between the digital and physical worlds. Recent advancements in deep learning and neural networks have propelled computer vision beyond human capabilities in tasks such as object detection and labelling.
This technology is employed in self-driving cars, enabling them to perceive their surroundings. It plays a pivotal role in facial recognition, matching images of people's faces to their identities. Additionally, it's crucial for augmented and mixed reality applications, allowing digital objects to overlay and interact with real-world imagery on devices like smartphones, tablets, and smart glasses. Online photo libraries such as Google Photos employ computer vision to detect objects and categorize images based on their content.
It enables these machines to not only perceive but also comprehend visual information, allowing them to take actions or provide recommendations based on what they "see."
In essence, computer vision mirrors human vision but faces the challenge of a shorter learning curve. Humans benefit from a lifetime of context and experience in distinguishing objects, gauging distances, detecting motion, and identifying anomalies in images. Computer vision, on the other hand, must rapidly acquire these capabilities through the utilization of cameras, vast datasets, and specialized algorithms.
One notable advantage of computer vision lies in its speed and precision. Systems trained for tasks such as product inspection or monitoring production processes can analyse thousands of items or actions per minute, detecting even minuscule defects or anomalies that might escape human notice. This capacity enables computer vision to swiftly surpass human capabilities in tasks requiring rapid and thorough visual analysis.
Another important concept in AI is cognitive computing.
Cognitive computing is a branch of artificial intelligence (AI) that aims to replicate and enhance the human thought process within a computer model. This approach aims to create systems that can interact with humans more intentionally and naturally by comprehending and interpreting human language and the meaning behind images. Cognitive computing excels at handling complex tasks such as natural language processing, data classification, knowledge management, and more. These systems are adept at performing advanced functions that involve human-like reasoning techniques. Their expertise includes extracting meaningful relationships from unstructured data, converting speech to text and vice versa, recognizing patterns in data, interpreting visual information, and utilizing machine learning to continually improve their performance. A key factor contributing to their success is their capacity to analyse and process vast amounts of data, which empowers them to derive fresh insights and anticipate solutions for a wide range of problems. This ability to continuously crunch numbers enhances their capabilities, making cognitive computing systems valuable across diverse fields, from business intelligence to scientific research and beyond.
Robotics
In its initial stages, robotics was a branch of mechanical engineering primarily focused on creating robots specialized in performing specific tasks like grasping or walking. These early robots relied solely on algorithmic control systems without any elements of artificial intelligence (AI).
As robots evolved and became more advanced, there arose a demand for smarter control mechanisms. This necessity led to the emergence of cognitive robotics research, which incorporated AI-based control structures into robotic systems.
In contemporary times, the fields of robotics and AI research extensively intersect, reflecting a substantial and vital synergy between the two disciplines.
Conclusion
AI works by receiving input data in various forms, analysing it using rules and algorithms, making predictions or taking actions, and producing an outcome.
The outcome of AI systems is evaluated through analysis, discovery, and feedback mechanisms.
Adjustments to input data, rules, algorithms, and target outcomes are made based on these evaluations, creating a continuous loop.
Five key components of AI are: Machine Learning, Deep Learning, Neural Network, Natural Language Processing, and Computer Vision.
Machine Learning is central to AI, allowing computers to learn from data and improve their performance over time.
Different methods for training AI models are explained, including Supervised Learning, Unsupervised Learning, Reinforcement Learning, Deep Learning, and Transfer Learning.
AI models like GPT-4 and Bard are examples of high-powered prediction machines, developed using large datasets and preprocessing.
AI models are deployed into applications to make predictions or decisions based on new, unseen data, where most users interact with them.
Machine Learning, while powerful, falls under the category of narrow AI, excelling in specific tasks but requiring fresh training for new purposes.
Opacity and unpredictability are significant characteristics of machine learning, making it challenging to fully understand the internal workings and predict the outputs of AI systems.
Machine learning systems require large training datasets and significant computational resources, often referred to as "big data" requirements.
Natural Language Processing (NLP) is a subfield of AI concerned with processing human languages, achieving human-like language understanding for various tasks and applications.
NLP can analyse texts in different languages, modes, and genres, as long as they are naturally occurring and not specifically constructed for analysis.
Computer vision, empowers computers and systems to extract meaningful insights from digital images, videos, and other visual inputs.
Cognitive computing seeks to replicate and enhance human thought processes within a computer model, facilitating natural interaction between humans and machines.
Cognitive computing excels in tasks like natural language processing, data mining, pattern recognition, and machine learning, thanks to its ability to analyse and process large amounts of data.