How does Artificial Intelligence actually learn? 

Key Takeaways

• • Ai learns from data, experience and feedback through a number of learning techniques.

• • The role of human actors is pivotal in shaping Ai’s trajectory.

• • Successful Ai implementation demands the ongoing collaboration of both technical and non-technical experts to manage risks and guide Ai toward safe, legal and ethical outcomes.

Introduction

In today’s interconnected world, billions of individuals and organisations engage on the internet on a daily basis. We post, take photos, watch content, shop, conduct court, transact and engage in many other digitally mediated activities. Business functions too are increasingly mediated through the internet, thereby contributing to the massive amount of data human beings generate each day. At the time of writing, human beings create 120 zettabytes each day. This phenomenon has given rise to the term “Big Data” – referring to the growing variety of data sources that arrives in increasing volumes and with more velocity.

The Value of Data

While this data is intrinsically valuable, its value must first be discovered. Our human minds needs a way to make sense of this sea of data. We need more powerful data processors to handle these various forms of data, that come in massive volumes, and at incredible speeds. Human brains simply do not have the computational capacity for this work. This is where Ai and ‘Machine Learning’ comes in.

Today, big data has become capital. the new oil. The most powerful corporations today are no longer those that own strong fixed capital assets. Rather, it is those in the intangible, tech economy, like Amazon, Google, Apple, Facebook, Open Ai etc.

Machine Learning

Machine learning is the practice of developing computer systems that can learn from data, experience, and feedback – without being explicitly programmed. Essentially, its is a type of Artificial Intelligence (Ai) that enables machines to learn and improve from experience, without human intervention. At its core, machine learning is about drawing inferences of hidden patterns, rules, factors, and correlations in data, by observing and processing problem-relevant data.

For example, imagine teaching a child to differentiate between cats and dogs. Instead of explaining the difference, you show them pictures of various cats and dogs. Over time, even without an explanation, the child learns to identify them. That’s how machine learning works with data. The value of machine learning, then, lies in the way that the output can be used for future tasks, such making predictions, planning, classification, recommendations, image recognition, natural language processing etc.

There are a number of techniques used to train Ai, including supervised learning, unsupervised learning, reinforcement learning, transfer learning, online learning, generative adversarial networks (GANs), data augmentation, active learning and evolutionary algorithms. In this peace, I will focus on the main techniques used to train Ai algorithms today.

Supervised learning

Supervised learning is a type of machine learning in which an algorithm learns to map an input to an output based on labeled examples provided in a training dataset. For example, lets say a security agency wants to streamline its security processes by developing an Ai system to detect potential fraudulent passports or IDs.

1. 1. Training DataThe security agencies will collect thousands of scanned images of both genuine and fraudulent passports and ID’s. The data is divided into training sets and testing sets (and sometimes even validation sets). Each image in the training set is labeled by human beings, either as “genuine” or “fraudulent.”

1. 1. Learning ProcessThe computer analyses the features of each passport image. It might notice that fraudulent ones tend to have certain inconsistencies in holograms, typeface, or layout for instance. Over time, and with enough labeled examples, it refines its own internal recipe to better differentiate between genuine and fraudulent passports.

1. 1. TestingAfter the training, new scanned images of passports (from testing data) are presented to the system. The system then predicts whether they are genuine or fraudulent based on its training.

Supervised Learning would be preferred when clear categories are present in the data, such as distinct safe and unsafe scenarios, allowing the algorithm to learn patterns and correlations from labeled examples.

Unsupervised learning

Unsupervised Learning, on the other hand, is valuable when we want to uncover new insights or when we are dealing with dynamic situations where patterns and challenges aren’t always well-defined or understood. The Ai system is given unlabelled data and is expected to find patterns and relationships on its own. For instance, lets imagine the local Police wants to train an Ai to analyse patterns in public gatherings for security threats. They might want to understand the dynamics of large public gatherings to preemptively identify potential security threats, but they don’t always know what they’re looking for.

1. 1. Training DataThe project team will collect various data from public events including CCTV footage, number of attendees, noise levels, online chatter leading up to the event, and so on. This data isn’t labeled—it’s just raw information without specific markers indicating a threat.

1. 1. Learning ProcessThe computer sifts through this data looking for underlying patterns. It might notice that certain spikes in online chatter correlate with increased agitation in certain parts of the gathering, or that certain patterns of movement frequently precede disruptions or violence.

1. 1. OutcomeThe computer can then cluster similar events or moments together. The police can review these clusters to understand different types of situations that tend to arise in public gatherings, helping them plan better for future events.

Reinforcement learning

Here the Ai system is trained through a trial-and-error process where it receives rewards or penalties based on its actions and learns to maximize rewards over time. It is particularly valuable in scenarios where an Ai algorithm needs to make sequential decisions in dynamic and uncertain environments. This approach is used when the optimal actions aren’t explicitly known, and the Ai algorithm learns through trial and error to maximise a reward signal.

For instance, consider training an Ai self-driving car to navigate through a complex and unfamiliar terrain, like central Joburg. The Ai needs to learn how to make decisions in real-time based on its actions and their outcomes.

1. 1. Training Data and EnvironmentThe project team first creates a simulation environment where the self-driving car can interact with the terrain. This environment provides the Ai algorithm with a way to perceive the current state (position and surroundings), take actions (move, turn, etc.), and receive feedback in the form of rewards or penalties based on the quality of its decisions. Unlike supervised learning, the Ai algorithm isn’t provided with explicit correct actions; it must explore and learn from its actions.

1. 1. Learning ProcessThe Ai agent starts with limited knowledge and explores the environment by taking actions and observing the outcomes. It learns from the received rewards to adjust its decision-making strategy over time. For instance, it might initially move randomly, but as it receives positive rewards for moving towards its goal and negative rewards for collisions, it learns to navigate more effectively.

1. 1. OutcomeThrough this iterative process, the Ai agent gradually improves its navigation skills. It learns which actions are more likely to lead it to its goal and avoids actions that result in negative outcomes. Over time, the Ai develops a policy—a set of rules guiding its decisions—that allows it to navigate complex terrains successfully, even ones it hasn’t encountered before. This learned policy can then be deployed to control robotic vehicles in various challenging environments, enabling them to make intelligent decisions based on their experiences.

Generative Adversarial Networks (GANs)

Generative Adversarial Networks (GANs) are a class of machine learning models used in unsupervised learning tasks, particularly in the field of Generative Ai. GANs consist of two neural networks, the generator and the discriminator, that are trained together through a competitive process, hence the term “adversarial.”

The generator creates fake data and the discriminator attempts to detect the fakes. This adversarial process continues back and fourth until the generator produces data that is indistinguishable from real data, at which point the GAN has achieved its goal. This technique is used to create image and video, including deepfakes, by producing synthetic data instances that share characteristics with the original training data.

Human-in-the-loop

In all these techniques, human beings play a central role through design choices and guidance that shape the trajectory of Ai’s learning. Whether it’s the curation of training data in supervised learning, or the intuitive formulation of reward structures in reinforcement learning, human expertise serves as the guiding compass for Ai.

💡 Humans are not just passive providers of data. We are active architects of Ai’s learning environment. We set the North Star by defining Ai’s problem objectives, and we also establish the yardsticks against which Ai models is measured. – Keketso Kgomosotho

As Ai continues to learn and evolve in an autonomous direction, the importance of human wisdom and insight will becomes increasingly paramount – to ensuring that Ai systems are trustworthy, and are aligned with legal, ethical, safety, and human rights considerations.

A symbiotic relationship

My sense is that humans and Ai will increasingly develop a symbiotic relationship, where Ai augments human capabilities while humans, in turn, steer Ai towards responsible and beneficial outcomes.

But in order to achieve the latter, technical experts and non-technical experts must come together across disciplines to set the North Star for Ai. That is, to ensure that ethical, legal, and societal risks and disruptions are anticipated and managed from conception, throughout the life cycle of Ai systems.

To effectively harness this potential, fostering collaboration between technical and non-technical experts is paramount. This collaboration ensures that AI’s trajectory aligns with ethical, legal, and societal considerations, safeguarding businesses from unforeseen risks and disruptions.

Conclusion

For businesses integrating Ai, this means that successful Ai implementation demands the ongoing collaboration of both technical and non-technical experts to manage risks and guide Ai toward safe, legal and ethical outcomes. This is no small feat. At TECHila Law, we embrace this challenge through a multi-disciplinary approach, merging technical and non-technical expertise into a cohesive offering to provide unique and comprehensive solutions that are alive to both the technical and non-technical dimensions of Ai.

 

Keketso Kgomosotho is an award winning Attorney and Researcher with over a decade of experience in Law, Artificial Intelligence, and Human Rights. He is an Ars Iuris PhD Fellow at the University of Vienna, where he is actively researching Ai governance from an African perspective, and he is a A member of the South African Ai Association (SAAIA). As co-founder of TECHila Law, his focus is on helping clients navigate the complex intersection of Ai, law and ethics while leveraging the power of Ai – making him a recognised authority in the field of Ai training, integration and governance. Keketso has a practical track record, having advised clients, including multinational corporations, on various aspects of data and technology governance. He’s been recognised as one of the Top 200 Young South Africans by the Mail & Guardian in the category of Law and Justice.