Open tgardner4 opened 5 years ago
ftmashari
commented
5 years ago Interesting! I have recently been reading about how machine learning is imposing irreproducibility to science (see https://www.bbc.com/news/science-environment-47267081). Hope we can also cover that in this session.
borisevichdi
commented
5 years ago Wow, that's a great topic!
Particularly, because I believe, that no, they should not. Because all the "trainings" and "workshops" does not properly teach people data science, but instead create a false sense of having expertise. While in fact, data science is a huge topic, that require permanent practice. And therefore, biologists think that they understand how to do proper reproducible analysis, while they know nothing ~Jon Snow~, and this false sense of security damages reproducibility even further. It would be so much better, IMO, if the biologists were to come to professional statisticians every time instead, and focused on their research topic; rather than they were to try to become a jack of all trades.
I'd be happy to join the debate on the "dark side" :)
The cure for scientific irreproducibility already exists. Why won't we embrace it?
For more than 10 years the scientific community has been grappling with the problems and impact of irreproducibility. But the institution of science is still largely business-as-usual...operating in a mode that is not sustainable. It is failing to deliver the societal benefits that $120B a year of global public investment and $200B a year of private investment should be delivering (see https://endpts.com/pharmas-broken-business-model-an-industry-on-the-brink-of-terminal-decline/).
Solutions to many of these problems already exist. They are the data-driven methods of Design of Experiments, measurement systems analysis, quality engineering, and data science. They've been put into practice over the past 70 years in manufacturing, supply chains, engineering, business operations. In biosciences, I've seen them eliminate experimental error, save discovery programs from absolute failure, and double the pace of discovery.
Yet, the institution of science is slow to adopt them and even dismissive of them. The methods aren't taught to freshman in college, or graduate students. They aren't practiced routinely in the lab. They aren't discussed, debated and elevated to a preeminent role in scientific discourse.
Business-as-usual in science is not a sustainable path. What is going to change it?
In this discussion I would like to uncover and debate what are the underlying drivers of scientific culture that prevent it from embracing the tools of experimental design, statistics and data science…and explore how to redirect that culture into a new era of computer-assisted scientific experimentation.