AI is reshaping scientific exploration at a rapid pace and it became really interesting, what kind of science we will see tomorrow? As the most dynamic and exciting opportunity, the latest AI algorithms are probing the evolution of galaxies, calculating quantum wave functions, discovering new chemical compounds and more… it’s only a short list of all developments, by the way. So, let’s take a closer look on this hot-topic nowadays and dive a little bit into thinking what to expect next.
Powerful Tool to Bring New Explanations and Predictions About the Universe
Scientists are arguing that the latest techniques in machine learning and AI represent a fundamentally new way of doing science. One such approach, known as generative modeling, can help identify the most plausible theory among competing explanations for observational data, based solely on the data, and, importantly, without any preprogrammed knowledge of what physical processes might be at work in the system under study.
In more detail, generative modeling copies the evolutionary approach of the natural world with cloud computing to provide thousands of solutions to a single engineering problem. Essentially, generative modeling asks how likely it is, given condition X, that you’ll observe outcome Y.
Trending AI Articles:
1. Data Science Simplified Part 1: Principles and Process
2. Getting Started with Building Realtime API Infrastructure
3. How I used machine learning as inspiration for physical paintings
4. Reinforcement Learning for Autonomous Vehicle Route Optimisation
The most recognizable generative modeling systems are GANs (generative adversarial networks) — the systems that can repair images that have damaged or missing pixels, or they can make blurry photographs sharp. And most importantly, Gan can make fake photos.
So what is the role of generative modeling for science? Well, one of the main usages is the opportunity to apply a relatively new approach to scientific research and the process of hypotheses generating. Let’s see what its creators said about it:
“It’s basically a third approach, between observation and simulation. It’s a different way to attack a problem.” — Kevin Schawinski, an astrophysicist and one of generative modeling’s most enthusiastic proponents.
Schawinski, by the way, working on his doctorate, he faced the task of classifying thousands of galaxies based on their appearance. Because no readily available software existed for the job, he decided to crowdsource it — and so the Galaxy Zoo citizen science project was born. Beginning in 2007, ordinary computer users helped astronomers by logging their best guesses as to which galaxy belonged in which category. Schawinski put it this way:
“Today, a talented scientist with a background in machine learning and access to cloud computing could do the whole thing in an afternoon.”
Amazing, isn’t it?
Living in a Data-Driven World AI Is a Must
AI helps humans to do routine tasks faster, but in the future, this assistance will be needed even more. What we have today? Everyone is dealing with an unprecedented amount of information. Nearly 90% of global data has been created only over the past few years! What is more striking, the amount of data in the world increased by nearly 10 times in every two years over the past three decades.
Machines, like people, store information in memory and eventually become smarter. But unlike us, they do not know things like short-term memory loss, overload of information, sleep disorder or inattention. That’s how machines can identify cancer, make classical symphonies, and eventually compete with humanity.
So what this is all about?
Of course, scientists have to operate with large data arrays every day. In fact, most of the work of scientists is routine tasks that take a lot of time. But AI can solve this problem, execute these operations, allowing scientists to focus on metadata and on more global scales of research. And yes, it’s not only about scientists, but it’s also about all possible occupations we have today.
More and More Breakthroughs in Medicine, Biology, and Chemistry
But come on, now, let’s move to all key breakthroughs AI made in science. I decided to start with medicine, biology, and chemistry since these areas have already been largely improved by the AI. So, for example, AI fundamentally changes the usual health care system and makes it more patient-oriented. I’m not going to talk too long about it, here are the main achievements that speak for themselves:
AI helps automate the drug discovery process: Identifying protein crystallization
A group of researchers from MARCO (MAchine Recognition of Crystallization Outcomes), together with the developers of Google, have developed a protein crystal recognition tool using convolutional neural networks. By analyzing protein crystals, one can recognize the structure of a complex molecule, thus determine its functions and subsequently create a drug aimed at it.
Chemical space exploration guided by deep neural networks
A group of researchers from the Skoltech Center CDISE (Center for Scientific and Engineering Computing Technologies for Problems with Large Data Arrays) and the Helmholtz Center in Munich created a neural network for visualizing the space of chemical compounds. The new method will help create new chemical compounds and analyze existing ones.
Scientists have used the method of reducing the dimension of t-SNE along with a deep neural network. Thus, the created neural network, represented as a multidimensional structure of the compound of interest, generates the coordinates of this compound in a two-dimensional form on a certain plane. In this case, molecules with similar properties are located nearby, which allows grouping the compounds into classes corresponding to a particular property.
In the future, scientists plan to create a number of programs that will allow researchers to view the distribution of new compounds relative to those already studied and described in the literature. This will help to quickly carry out the R & D phase of research when searching for new drugs.
Controlling an organic synthesis robot with machine learning to search for new reactivity
Researchers at the University of Glasgow have train an artificial neural network to find new compounds and determine chemical reactions. The system monitors in real time the substances that have arisen during the reaction and disappeared using infrared spectroscopy methods based on nuclear magnetic resonance. After that, the neural network, on the basis of which the system is developed, processes the received data and makes decisions on further reaction paths.
Modeling polypharmacy side effects with graph convolutional networks
Stanford University researchers have developed an artificial intelligence-driven system that can analyze the interaction of two drugs.
As a solution, scientists have created an extensive system of deep learning, based on data on the interaction of more than 19,000 proteins, and interactions with them of various drugs. The system is called Decagon, and it can effectively predict the interaction of any two different drugs.
At the moment, the system is analyzing the interaction of no more than two drugs, but scientists expect in the future to expand them to more complex drug combinations.
AI is getting closer to replacing animal testing
A group of researchers led by Thomas Hartung from Johns Hopkins University have developed a computerized approach to detect toxicity of chemicals, exceeding the accuracy of animal experiments. To create it, binary similarity coefficients were used, and the system itself was trained on information about more than 866 thousand hazard properties of chemicals common in experiments.
Scientists used data on 80 thousand chemicals from dossiers collected by the European Chemicals Agency: the final dataset contains information on 866 thousand properties and indications of toxicity and possible adverse effects of all chemicals analyzed. As a result, researchers were able to achieve the accuracy of toxicological analysis in 87 percent of cases: for comparison, a single experiment involving an animal has an accuracy of 57 percent, and a repeat one — 81 percent.
Neural network for the treatment of rare diseases
Engineers at the University of Toronto helped solve the problem of limited data for rare diseases for machine learning algorithms by simulating X-rays using competition between two neural networks. The new approach has improved the diagnosis of rare diseases by almost one and a half times.
Scientists have applied the technology of deep convolutional generative-contention network (DCGAN). Generative-contention network is a kind of algorithm based on the interaction of two networks. One generates images, and the other tries to distinguish artificial from real ones. Training takes place until the second machine stops noticing the differences. As soon as a sufficient number of X-ray images is created, a convolutional neural network is connected to their analysis, which tries to recognize signs of the disease on them.
Comparing the effectiveness of the new approach with the traditional one, the developers found that the accuracy of diagnosis of common diseases increased by 20%, and for some rare ones — by 40%.
The Look into the Future of Science and My Opinion…
Science now is not only the concerted human but also machine effort to understand or to understand better, the history of the natural world and how the natural world works.
Maybe it sounds paradoxical, but the future with AI will be something common for us and completely invisible. No machine uprising. Every time when any loud technical innovation comes to our life, it causes a lot of emotions, and after a while everything becomes normal.
Most likely, AI can answer any question concerning science. But, we can not understand where did this answer come from.
How can it be? Here’s an example: let’s imagine that someone throws a ball at a person and he catches it. The person can predict the trajectory of the flight of the ball unconsciously ( he or she doesn’t make any calculations for this). On the other hand, the trajectory of the ball can be determined not by using the human neural network, but by the methods of classical physics, that is, by calculating the equation.
Just as a person predicts the trajectory of the ball, so in the future it will be possible to predict the development of any process in the universe using artificial intelligence. But such a prediction will not look like physical theory in general understanding.
What are your views on this?
If you enjoyed this post, leave a comment below, and of course, don’t forget about a few claps!
Explore more useful tips and read dope topics on AI, ML & Data Science from my Medium or Instagram blog.