We have written dozens of articles about generative AI already, covering all possible angles, but just realized that we have never covered the basics, and explained what is generative AI in the first place.
Generative AI refers to a category of AI algorithms that look for patterns and structures in the sample data and come up with new ones. For example, it can simulate discussions and learn to find out how we, people, would be satisfied with the results. But it does it billions of times a day. So it improves at an unbelievable rate.
This kind of AI can be used for a wide variety of tasks, such as creating illustrations, writing text, composing music, or generating new ideas.
Depending on the task, generative algorithms are trained and work in different ways. There are many models and methods, but we’ll most likely regularly meet with two these days:
- Generative Adversarial Networks (GANs), for image or sound generation, think of Midjourney, Blue Willow, AIVA or Loud.ly.
- And Transformers, that are behind text generators, like ChatGPT or Bard – also known as Large Language Models (or LLMs).
(The Transformers-LLM relation is like the insects-beetles situation. Transformers are a type of LLM, but not all LLMs are transformers. Transformers are the most popular architecture for LLMs these days because they are very effective for natural language processing tasks, but they are not the only ones.)
We’ve already introduced how GANs work, you can easily imagine it through the forger-art detective analogy. Transformers on the other hand can be best grasped as brilliant autocomplete machines: they are trained to predict what comes next based on the previously given data.
How do they learn?
There are many methods to train generative AI models, and most of the now successful models use a combination of various methods.
GANs are typically taught through unsupervised learning – which means that the AI is not learning from labelled data – there is no one showing it pictures labelled as cat and not-cat. The “supervision” in GANs comes from the adversarial process where the generator network (the forger) tries to fool the discriminator network (the art detective), and the discriminator network tries to correctly classify whether an input is real or fake. For a simpler explanation, head over to the previously linked GAN article.
Large language models, like ChatGPT or Bard are often trained using both supervised and self-supervised learning methods.
For instance, the GPT (Generative Pretrained Transformer) series by OpenAI is trained using a form of self-supervised learning – according to ChatGPT. The models are given a large corpus of text and learn to predict the next word in a sentence given the previous words. In this case, the “labels” are the subsequent words in the sentences themselves, so no explicit labels are needed.
Google’s Bard on the other hand was educated through a combination of supervised and self-supervised learning, on a massive dataset of text and code, including both labeled and unlabeled data, using the PaLM 2 model.
ChatGPT was the first text-generating AI algorithm that became a massive hit among everyday people, getting a hundred million users in just two months, making it the fastest-growing consumer application in history. Since then there are several other widely successful algorithms, like Google’s Bard and Microsoft’s Bing – the latter is also based on GPT-4 and was customised for search.
These algorithms were not specifically trained for medical purposes and the regulations are lacking, but they can still be used for a number of purposes even in medicine, such as writing emails or summarising research papers.
Based on the first published studies, three main areas of focus for ChatGPT emerged, namely 1. clinical use, 2. answering medical questions and assisting in education, and 3. scientific writing and research.
There are many open questions, but we already have some answers as well, such as LLMs are not becoming authors in scientific papers, and that medical alternatives will arrive soon.
Large language models specifically trained for medicine
Given the potential price of mistakes in healthcare, using just any LLM is a no-go for many tasks. You don’t want a diagnostic aid that generates answers from thin air every now and then. Here is where specialised algorithms enter the arena.
As of now, the most advanced contestant is Google’s Med-PaLM 2. We introduced it in a detailed post earlier and discussed why getting help from a medical chatbot as the first line of contact might soon become a lower-risk option than waiting for a doctor.
A lot has happened in the past 6 months, and Google recently announced how much their medical LLM improved.
A new study published in Nature revealed that Med-PaLM 2 answered medically related questions with 92.6% accuracy – mobihealthnews reported. Although the detailed data shows that MedPaLM still has areas in which it needs to improve, the results are nevertheless impressive.
Generative AI in practice
Medical LLMs are not yet widely available, Med-PaLM will be released for a select group of users soon and is also reported to already be in use by Mayo Clinics. However, many generative AI tools are available for everyone and can bring practical benefits to healthcare professionals. Let’s see a few existing use cases!
Generative AI can also be used to come up with new ideas or inspirations for design processes, from suggesting brainstorming techniques to listing ideas. We also used it to design digital health technologies – an interesting, but not very fruitful venture.
Digital avatars also belong to the realm of generative AI, The Medical Futurist already has a deep fake twin, also featured in this video. These avatars are yet too uncanny to use in patient-facing roles – anyone would find a written note from the doctor more assuring than a message delivered by an unblinking synthetic human.
On the other hand, such video-doctors could tirelessly deliver professional training to colleagues, instead of someone spending hundreds of hours in front of cameras for recording the same material.
Many free AI tools are just as easy to use as an app on our phones or laptops – it just needs a little practice. We encourage every healthcare professional to play with these algorithms to become friends before they arrive at the office. Giving a little time to master our prompt engineering skills – how to ask questions efficiently – will deliver a massive payback in the future.
The Medical Futurist summarized what we can expect from AI in healthcare in the coming period. On top of the list is the birth of proper regulatory frameworks, specialty-based AI recommendations and guidelines, and seeing yet undiscovered medical areas where AI can assist us. For a thorough but gentle introduction to the AI era, I recommend our Introduction to Artificial Intelligence In Medicine And Healthcare course, which was recently updated with a bonus chapter.
These skills are slowly becoming more and more necessary. Multi-modal LLMs – algorithms that will be able to analyse not just text, but all content formats, from images to sounds to videos – will arrive. We need to find our creativity in this new world, and without a working knowledge, both creative and cognitive professionals will be left behind.
The post Generative AI Explained – Its Impact And Future In Healthcare appeared first on The Medical Futurist.