The Second Serbian International Conference on Applied Artificial Intelligence (SICAAI).

Abstract

Recently, ChatGPT (Generative pre-training transformer) is often mentioned as a tool for generating real text, i.e. text that, without a deeper analysis, is very difficult to distinguish from the one written by a human. Generative models of this type can also generate other forms of data – such as audio recordings and images. When generating images, there are several models available, such as Dall-E 2 (offered by OpenAI – authors of the ChatGPT model), Wombo Dream, Midjourney, and many others. These models are accessed through a web browser, usually with restrictions and/or payment per generated image, while there are also models that can be run on your own architectures – such as Big Sleep AI developed in Python, which is not only freely available, its source code is already freely available. Such models are based on the so-called generative deep convolutional artificial neural networks (DCANN). Such methods are historically and currently in use for image classification purposes. An image is fed to the “input” of the neural network, which passes through the layers of the network in which convolution operations take place with filters whose values are adjusted, and at the output they give the intended class of the image. Generative models, like the ones mentioned earlier, work in an inverse way – where based on the input of the class of the desired image, an image is generated from a field of random values. Examples of such applications exist from before, such as “On urinary bladder cancer diagnosis: Utilization of deep convolutional generative adversarial networks for data augmentation”, which, based on the numerical input of class values, can generate an image representing healthy or diseased tissue. Generative models like this one are proving to be an important tool in medicine-focussed machine learning model training, as obtaining a dataset that is both numerous (thousands of images), balanced (different image classes, e.g. healthy vs. diseased tissue), and high-quality can be almost impossible. This is especially notable when rare diseases are discussed. In cases where a comparatively poor dataset is provided, it can be augmented using such advanced image generation techniques. For modern image generators, users want to enable simpler input, so generation models are combined with an additional artificial intelligence model for the interpretation of “spoken” language, which enables text input (e.g. “yellow submarine”, “dog running through a forest”) instead numerical representations of the class. Models combined in this way are called CLIP (contrastive language-image pretraining) models. Since the models are “general”, i.e. they are expected to be able to generate a large number of different subjects in the images, they require a large amount of images (e.g. Dall-E 2 is trained with more than 400 million image-text pairs), which are retrieved from the Internet, where tags, titles and descriptions are taken along with the image as text. This approach of obtaining a vast amount of images automatically is significantly different from a careful, curated process of image collection for the model training today – in which images are checked by experts in the given field to assure their quality. The question arises of the need for such a curated process in data collection, especially field-wise. Should images used for model training in fields like medicine be carefully validated, while this is unnecessary for different fields of application? Although this is the only realistic way to collect such a large number of images, many artists raise the ethical question of the right to use their images – that is, are they actually responsible for the images generated by the artificial intelligence model, since their images are used to train these networks?