The rise of artifitial intelligence and its applications in new generation science

With the fast development of computational resources, we’ve seen the rise of the AI (artifitial intelligence, or neural networks) in various real world applications, including object recognition in self-driving cars, facial recognition for automated security, text generation and many other. As I’m in science and engineering field, the exciting thing about AI for me is its applications in scientific research, which creates a great amount of opportunities in terms of research methodology and philosophy.

Traditional scientific research can be described as such two trends:

• Starting from experiment, observe the experimental result, guess a theory, deduce new facts from the developed theory, perform a new experiment to verify the result, confirm the theory (fundamental science)
• With the matual theory combined with real world demands, develop a system that is useful in certain applications (application and engineering)

AI can bring new light into both of the two approaches. For fundamental science research, the neural networks can act as a powerful tool for data analysis, for example, the reduced space can help scientists capture the essential physics of a complicated phenomenon. For application and engineering, AI lets one ignore the “true” theory behind the phenomenon but directly find new implications and applications.

That said, the AI isn’t omnipotent. Though there has been works on semi- or unsupervised machine learning, the current AI still needs great amount of data to work, where essentially the problem becomes a mathematical fitting problem. This impose some limitation on the application of AI in science, where data might be hard to acquire. On the other hand, ignoring the inner physics of a system also limits its application in engineering field. Each trained AI system can only be used for a specific task, and for a different task a new neural network has to be retrained. But the trained system is generally good at that specific task. One can understand this as the AI solves a subset of a larger problem, thus it effectively makes some “intelligent” approximation.

AI can definitely help the development of our science and engineering, but such new techonology hasn’t yet been fully explored in science community. I’m sure it will be a necessary tool to learn for everyone in the community.