The AI landscape

Computer Science

Computer science is the study of computation, information, and automation

Wikipedia

Computer science encompasses AI as we define it, as computers will be the most practical tool for its implementation.

Artificial Intelligence

In the course Elements of AI, from the University of Helsinki, AI revolves around the following two concepts.

Autonomy The ability to perform tasks in complex environments without constant guidance by a user.

Adaptivity The ability to improve performance by learning from experience.

AI can be seen as a domain of computer science aiming at studying artificial adaptive systems capable of performing autonomously.

Data Science

Data science combines math and statistics, specialized programming, advanced analytics, artificial intelligence (AI) and machine learning with specific subject matter expertise to uncover actionable insights hidden in an organization's data. These insights can be used to guide decision making and strategic planning.

IBM website

You may ask yourself: what does data science have to do with all this? It is introduced, together with AI, as one of the parent domains of machine learning.

Machine Learning

Machine learning is the subset of artificial intelligence (AI) focused on algorithms that can "learn" the patterns of training data and, subsequently, make accurate inferences about new data. This pattern recognition ability enables machine learning models to make decisions or predictions without explicit, hard-coded instructions.

IBM website

You may have expected the word data science to pop up in the definition. The key seems to be the inferences made from the data.

Deep Learning

When we bring multi-layered neural networks to machine learning, we get a subset called deep learning.

Mapping it

Machine learning being the intersection of AI and data science is arguable. But as said before, the boundaries are blurry and it is practical enough as an introduction and reminder that machine learning has a place in data science.

Three basic types of problems in machine learning

Now we will see three common problems we try to solve with machine learning. We can find them as building blocks of more complex problems.

Classification (Supervised)

Predicting a discrete label or category for a given input.

• Goal: Assign an element to one of several predefined classes.

• Examples: Spam vs. non-spam detection, handwritten digit recognition (0 to 9).

Regression (Supervised)

Predicting a continuous numerical value for a given input.

• Goal: Estimate a quantity along a continuous scale.

• Examples: Predicting house prices based on features, forecasting tomorrow's temperature.

Clustering (Unsupervised)

Discovering natural groupings or patterns in unlabelled data.

• Goal: Group similar data points together without predefined labels.

• Examples: Customer segmentation for marketing.

Generating outputs, as we will see in GenAI, seems at first disconnected from those three problems, but when diving into the building blocks, some of them reappear. Now let's have a look at deep learning and generative architectures.

Diving deeper towards GenAI

Training objectives

Before looking at how these networks are built, it helps to know what we want from them. In machine learning, several kinds of models exist. One family of models is called discriminative: from what they learnt, they determine to which class an input belongs. Another one is generative: from what they learnt in the training phase, they generate data that fits the training.

Here are three generative training objectives.

Next-token prediction

Let's take sentence generation. One way to generate a sentence is to pick words or even characters one after the other. Once trained, we want the model to be able to guess a plausible token at each step. During training, we show the model a piece of text, hide the word that comes next, and ask it to guess. Repeat billions of times. Nothing else is needed: the model is never told what grammar is, or what a fact is. This is how LLMs and SLMs are trained.

Choosing one token out of a finite vocabulary. Does it remind you of one of the three basic types of problems seen in machine learning?

Diffusion

An output is generated out of noise. At each step, a network takes a noisy image and predicts a slightly less noisy one, and after enough steps an image appears. Any architecture that maps an image to an image of the same size can do this, U-Net historically, or more recently DiT.

At each step, the model predicts a continuous value for every pixel: the noise to remove. Predicting a continuous value through supervised learning. Does it ring a bell?

GANs

In generative adversarial networks, two neural networks compete against each other, a generator and a discriminator. The generator (surprisingly) generates output candidates judged by the discriminator until it agrees.

Funnily enough, two of the three basic problems were hiding in there: classification for the next token, regression for the noise. In GANs, the discriminator is a plain binary classifier, whereas the generator has no target to predict. What makes these systems generative is not the basic block itself, but the steps and loops built around it. Now let's see how things are done.

Under the hood

Artificial neural networks

This concept is central to deep learning. A neural network is made up of several layers of neurons.

Giving an input to the network equates to giving values, called activations, to the first (input) layer. The goal is then to compute the activations of the last (output) layer from the network configuration.

Each neuron of a layer can be linked to neurons of the previous layer via a value called a weight. The value of a neuron is computed in two steps:

• A linear combination of the activations it is linked to on the previous layer, each multiplied by the weight of its link, plus a value attached to the neuron itself called a bias.

• An activation function, which takes that result as input and produces the neuron's activation. This function is usually non-linear.

Note: Weights and biases are fixed when computing an output; activations are what get computed. Training reverses this: activations are the by-product, weights and biases are what change.

In the training phase, when the model learns, the outputs are compared to the expected values through a cost function, which quantifies the error. We then minimise the cost by adjusting weights and biases.

Building blocks

Often, several consecutive layers of neurons perform similar characteristic operations. This leads us to look at blocks, i.e. a pattern of repeated layers.

Convolution blocks

In convolutional neural networks (CNNs), convolution layers form the first stage of the network. A typical input is an image. The idea is to capture the geometry of an area of the image, via a small array of learned weights called a filter, and then sweep the image from there. Each position produces one value, and together these values form a new image called a feature map. Layer after layer, the complexity of the image is captured.

Transformer blocks

One of the main traits of the transformer is to be able to capture the relationship between several parts of its inputs, typically a sentence. This is done via a mechanism called "attention".

Architectures

CNNs

Simple CNNs are built from convolution blocks. Their main advantages for image processing are:

• Local connectivity and weight sharing: each neuron looks at a small patch, and the same filter is reused across the whole image. Together they reduce the likelihood of overfitting, and drastically cut the number of parameters, which means fewer examples are needed.

• Translation equivariance: shift the input, the output shifts the same way, so a pattern is detected wherever it appears.

• Hierarchical feature learning: the first layers pick up edges and gradients, the next ones textures and simple shapes, the deepest ones full objects. This hierarchy emerges from training, which is what makes CNNs well suited to transfer learning.

Note: A convolution layer can only express a narrow set of functions compared to a dense one. It works because the restrictions match reality in natural images: neighbouring pixels really are related, and a pattern really does mean the same thing wherever it appears.

U-Net is an evolved type of neural network using convolution blocks. Image characteristics are extracted in a contraction phase, then the image is reconstructed during an expansion phase. These techniques have the specificity of linking what is identified to the place it is identified on the image, for example a tumour and its location on a medical image. This is also the shape a diffusion model needs, which is why U-Net ended up there.

Transformer architectures for LLMs and SLMs

In LLMs and SLMs, transformer blocks sit between the transformation of the input sentence into a matrix of vectors and the transformation of a distribution of likelihood over the vocabulary into a selection of the next token.

Placing modern GenAIs

GenAI is not new but you can find recent generative models in Natural Language Processing and Computer Vision.

Conclusion

Congratulations for reaching this part af the article. Now you have a rough idea of where the recent advance in GenAI sits and the relation between AI domains. You know as well what problems are solved and some ways to do it.