 In this video, I will explain why generative models are so important. We will show how generative models can help scientists come up with new hypotheses and achieve new and faster discoveries. Hi, I'm John Smith, IBM Fellow and Lead for Discovery Technology Foundations at IBM Research. OK, first some basics. Science is how we study and understand the world. Scientific discovery is a key process of inquiry in science that is built on generating and testing ideas. The scientific method is how we carry out the steps of scientific discovery. So let's explore this further. OK, first, every discovery using the scientific method starts with a question. As an example, let's consider the question, can we design a drink that astronauts can take to space? This can involve studying the question where the goal is to investigate prior knowledge, which in this case might be to research different ingredients, their nutritional properties, behaviors under different conditions, and so on. The goal of this studying is to build a foundation of knowledge in a context for hypothesizing about possible answers to the question. Next is where the creativity comes in. The next step is to come up with this hypothesis. And the hypothesis is a creative guess at the answer to the question. And a good hypothesis is one that can be refuted or proven wrong through testing and experimentation. The step of coming up with the hypothesis is often referred to as the eureka moment. It's when the scientist says, aha, in the case of the space drink, it might be proposing some novel powder-based formulation. In this case, the hypothesized answer to the space drink question can then be tested by making it and subjecting it to space conditions or testing it with astronauts. And this can lead not only to a solution to the question, but it can also be used to help generate new knowledge by assessing the results of that test. And then by reporting what was learned back to the scientific community. And this can have great value for leading to new questions or new answers to existing questions. So why is it so difficult to come up with a hypothesis in practice? The reason is that there are a very large number of possible answers to a scientific question. And the scientist may be too overwhelmed to come up with good answers, particularly if it's a really tough question. As an example, let's consider the case of a scientist wanting to design or discover a chemical that can effectively catalyze a reaction to reduce CO2 emissions. This is a question many scientists want to answer. However, the chemical space is practically infinite. There is no way to search over all possible chemicals to come up with an answer. Here, creativity is needed in generating ideas. Okay, so let's consider another example. Let's say we have the challenge of designing or discovering a small molecule drug that can treat Alzheimer's. Well, it's been estimated that the number of possible drug-like molecules is 10 to the 63. There's no way to brute force search over all possible drug-like molecules and test them to see what works. This is where the generative models come in. Okay, so let's see how generative models work. So generative models are a type of AI model that essentially learns from training data how to be novel. Generative models are different from other discriminative type models that are commonly used in machine learning. So let's look at discriminative models and how they work. Discriminative models are essentially given training data, and what they learn to do is to distinguish between different classes of that training data and effectively learn a decision boundary that can tell the difference between them. Now, on the other hand, generative models work completely differently. Generative models are also given training data. However, what the generative models learn to do is to model the distribution of that training data in such a way that they can generate new instances that belong to that distribution. Okay, so in the case of scientific discovery or in an application of molecule design or materials design, how can these generative models help? Well, it's possible to train these generative models from known molecules. So given these known molecules and their properties, it can be possible for these generative models to then propose new candidates that might fit a set of criteria that can answer a scientific question for an application. If we come back to the application in materials discovery, it can cost between $10 million to $100 million and take up to 10 years to design and discover a new material. Generative models can greatly shorten the time that it takes and reduce the cost. And this has application in so many different areas. It can help accelerate discovery for problems related to climate and sustainability, for therapeutics, and a broad set of problems in general. Thanks for watching. If you enjoyed this content, please like and subscribe or visit our website and download our toolkit GT4SD.