For example, when company Z needed to predict which products would be of most interest to its customers, it used a supervised learning model which, fed with previous purchase data, generated personalised recommendations according to customer preferences. In addition, it took advantage of the unsupervised learning technique to segment audiences, finding patterns of behaviour that are not so obvious at first, such as those users who prefer to shop at weekends or in the evening after finishing their activities, allowing specific and more targeted offers to be launched at those times. If we also look at the area of customer service, recurrent neural networks (RNN) have been implemented in chatbots, achieving more natural and fluid responses, improving interaction with users.
But the needs of company Z go beyond making product recommendations or personalising the service for its users. The reality is that the opportunities for applying AI are as diverse as the different needs that company Z may have. From the automatic generation of content and images for the catalogue, to the simulation and projection of future behaviours or the creation of product descriptions with hardly any human intervention. This is where generative models come into play, offering significant differential value.
What are Generative Models?
Unlike other types of AI models that are based on making classifications or predictions, generative models go a little further. These models not only process and analyse data, but are also capable of creating new data similar to the original. This is where the discussion of how creative they can be opens up, but it is clear that being able to generate ‘new’ content similar to the previous one is a major breakthrough. In company Z, this could mean generating images of products, simulating customer behaviour or even proposing product descriptions that are even closer to the preferences of their audience.
There are different types of generative models, but today I will focus on two of the most important: Generative Adversarial Networks (GAN) and Autoencoders, each with its own scope and applications.
GAN vs. Autoencoders
To understand how they work, let’s continue with a practical example. Imagine that company Z wants to create images of new products without having to hire a photographer. GANs may be the perfect tool for this.
Generative Adversarial Networks (GANs) work through a kind of competition between two neural networks: the generator and the discriminator. The generator creates false images (for example, a T-shirt that has never existed), while the discriminator evaluates whether the image is real or generated. Through this ‘game’, the generator continuously improves until it manages to make the images created almost indistinguishable from the real ones by the discriminator, it is a process of continuous improvement that happens very quickly until the generator reaches its maximum level.
On the other hand, autoencoders work differently. Returning to company Z, let’s now imagine that it wants to reduce the size of the images without losing too much quality. Autoencoders compress the information into a reduced ‘code’ and then reconstruct it, maintaining the essence of the original image. However, a special type of autoencoder, called a Variational Autoencoder (VAE), takes this to the next level, allowing not only the reconstruction of the image, but also the generation of new images from that reduced space.
How do GANs work in practice?
Let’s suppose that, based on current trends, company Z wants to generate images of T-shirts. The generator starts by creating random images, while the discriminator (which has seen thousands of real images) evaluates how realistic those images are. At first, the images may look like scribbles, but with time and many iterations, GANs can produce surprisingly realistic images.
In the financial sector, one example is using GANs to detect financial fraud. Instead of using sensitive customer data, banks can generate similar synthetic data, allowing them to test their fraud detection models without compromising user privacy. This same principle could be applied at Company Z to generate test data, optimising its processes without taking unnecessary risks.
And what about Variational Autoencoders (VAE)?
Unlike a traditional autoencoder, which focuses on compressing and reconstructing data, VAE allows company Z to explore new possibilities based on that same data. For example, company Z could generate product descriptions based on the most effective writing styles previously used. VAEs create a ‘map’ of possibilities, allowing not only the recreation of data, but also the generation of new creative variants.
Summary
In short, generative models open up a world of possibilities for Company Z. From generating images of new products to creating test data for simulations, or even innovating in the way they interact with customers, these models allow them not only to adapt and keep riding this new wave of AI, but also to anticipate what the market demands. Let’s continue discovering together how artificial intelligence can transform industries, always with company Z as our innovation laboratory.
