The number of research papers that mention AI or ML in the title and abstract is taking off.
The journal Nature asked 1600 researchers around the world how they use AI, and about their fears and hopes.
The most positive impact of AI was time:
ML provides a faster way to process data (76%),
it speeds up computation (58%),
it saves researchers time or money (54%),
it automates data acquisition (48%), and
it provides a faster way to write code (43%).
But also researchers value ML/AI for enabling research:
ML makes it possible to process new kinds of data (47%),
answers questions that are otherwise very difficult to solve (42%),
makes new discoveries (37%), and
generates new research hypotheses (33%).
There are also concerns about bias, fraud, power imbalances, energy use, and lack of reproducibility.
But by far the biggest fear was that ML and AI lead to more reliance on pattern recognition without understanding (68%).
Science-ready machine learning
Machine learning can fit complex functions for problems we have a hard time coming up with human-understandable rules. We give up understanding. We get flexibility.
But there’s a difference between bare-bones machine learning and a more holistic approach. We already have a lot of tools that make machine learning ready for science, they are just scattered all over the place:
Interpretable machine learning
Causal inference
Uncertainty quantification
Robustness
Ways to infuse domain knowledge
Representative data
Timo Freiesleben and I are writing the book “Supervised Machine Learning in Science”.
The current draft covers the justification for machine learning in science and the first chapters on domain knowledge and causal inference. We are looking for feedback on whether the book is on the right track.
Interested in being a beta reader?
👉Send an e-mail to readers@christophmolnar.com
We still experiment with the cover, but it’s gonna be a raven mascot this time!
Definitely would love to be a beta reader! I'll be writing you ;)