A leader's guide to GenAI technicals - What is AI

Artificial intelligence is on the tongues of many. Perhaps the most overused set of words at the moment.

Generative AI pulls the bandwagon.

We talked a lot about use cases, tests, or performance.

Here, we'll examine the basic technical concepts so that terms like neural networks, parameters, transformers, or embeddings are clear and understandable.

We'll also try to do it most efficiently. No mumbo jumbo.

Let's call it a fast-track course.

Welcome to episode one of a leader's guide to AI technicals.

Key Takeaways:

- Artificial Intelligence is the ability of machines to perform tasks typically requiring human intelligence.

- Most of the AI we use today is sophisticated Machine Learning models. It revolves around complex algorithms capable of learning and making decisions. It's based on data, parameters, and predefined objectives.

- There are four main types of ML: supervised learning, where models learn from labeled data; unsupervised learning, which finds patterns in unlabeled data; reinforcement learning, focusing on reward and penalty mechanisms; and self-supervised learning, a hybrid approach that allows models to generate their own supervisory signals from the data.

What is AI?

Before we move to Generative AI, let's start with the fundamental question.

It will help us later.

The definition of AI varies, but generally, it refers to giving the machine the ability to emulate human tasks.

In this case, any rule-based system that states, " Do this if this happens, else do that" is AI.

You might remember choosing enemies in strategy video games. They were labeled as AI and had different difficulty levels. In fact, these systems were giant trees of actions and reactions.

Nonetheless, they gave the computer the power to compete with human players on equal terms.

Fulfilling the definition.

Machine Learning - THE AI

ML and AI nowadays mean basically the same.

The vast majority of AI systems we hear about are actually sophisticated and complex ML models.

Although unexplainable (due to their sheer size and number of computations), they check the list of ML model's elements:

• Data is the information used to train, validate, and test a model. Depending on the model's purpose, it can be anything from numbers and text to images and sounds.
• Parameters are parts of the model that are learned from the training data. Think of them as the knobs that the model can adjust to make better predictions or decisions.
• Learning algorithm is a process that a machine learning model uses to learn from the data. The learning algorithm determines how the model adjusts its parameters to minimize prediction errors or improve performance on some tasks.
• Task is what the model is built to do. This might be recognizing images, predicting stock prices, translating languages, or recommending products.
• Objective is often framed as an objective or loss function. It's a mathematical way to measure how well the model performs its task.

Types of ML

Before we move to Generative AI, let's spend a minute on how we classify ML models.

We use "supervisory signals" to separate certain types of ML. These are the ways we can train a model.

Let's roll back a bit to frame it better. We have the data, parameters, algorithm, task, and objective.

Imagine you're in target practice. You have a bow and an arrow.

Your goal is to hit the bullseye. This is the task.

Your precision and accuracy are objectives that indicate how well you're doing.

Arrows hitting the target are your supervisory signals.

You use them as a reference to adjust parameters such as aim, power, stability, etc., to do better.

Supervisory signals are the key to learning.

These are so important in machine learning that we use them to classify ML systems. Let's look at the classification.

Supervised learning

This type of training is done on fully labeled data. This means that for each data point, there's a clear answer for the "right or wrong prediction" case.

The AI has to classify the image as a cat or dog. The training dataset contains images labeled as "cat" or "dog."

It sees a cat and says "cat." Good.

It sees a dog and says "cat." Bad.

Of course, actual business cases here are much more complex (e.g., spam filtering or sales prediction), but the underlying idea is the same.

Unsupervised learning

The dataset does not have labels, meaning no supervisory signal. This might seem naive, but the goal is to uncover unknown patterns in the data.

Clustering is a good example here. Imagine you have a big dataset of plant images. Unsupervised learning might be a good option for grouping them based on certain similarities.

Anomaly detection is also one of the most popular cases, as it finds a certain minority of elements that don't match the rest of the data points.

Reinforcement learning

This is a very specific type of training with a reward and penalty model. In this type of training, the AI agent performs actions in a specified environment to achieve a goal, just like in a game.

Each action is either rewarded or penalized. The model aims to get as much reward as possible by learning how to perform the best possible actions.

It is used, for example, in training robots to move and lift objects through a "trial and error" process.

The most commonly used examples here are GO and chess bots, which play games against themselves to develop the best possible strategies.

Self-supervised learning

It's a mixture of the first two. The data provides the supervision here. It's a very clever approach: the model is trained on a dataset, sees only a part of the input data, and predicts the rest.

Large language models are a perfect example of self-supervised learning.

The model is so complex, and the dataset is so big that there would be no other way to train it.

Summary

Each of the presented paradigms is unique and useful in various problems.

Choosing the correct method is the "live or die" of an AI system.

In the following chapter, we will focus on simple explanations of neural networks and deep learning so we can naturally move to Large Language Models.

Stay tuned!

For more ML and AI insights, subscribe or follow Sparkbit on LinkedIn.

If you're looking to start an AI project, you can book a free consultation with our CTO here: https://calendly.com/jedrek_sparkbit/ai-consultation

Author: Kornel Kania , AI Delivery Consultant at Sparkbit