Machine learning may be causing a reproducibility crisis in science

Use of machine learning techniques has been identified as a possible cause of reproducibilty problems affecting scientific research worldwide.

Artificial intelligence and machine learning are rapidly transforming how we analyse data, but these tools may be exacerbating the current reproducibility crisis in scientific research. Machine learning can be used to turn large data sets such as images from astronomical telescopes or pictures of cells into a system that can accurately classify new data, but exactly how it has arrived at a particular conclusion is difficult to determine.

This has led to many scientists taking a black-box approach to analysing data using machine learning algorithms, feeding the system their data sets and then using the resulting model to draw conclusions. The problem is that when new researchers try to replicate the research using different data sets, they often come out with different conclusions.

Part of the problem is that there isn't sufficient understanding of the limits of AI and machine learning. These systems can be incredibly useful for automating analysis of large data sets, but they require considerably more data points than a human researcher would typically need to begin drawing reliable conclusions. They're also very sensitive to finding patterns that may only exist in one particular data set, and will need to be tested with multiple sources to eliminate that issue.

BBC Science correspondent Pallab Ghosh recently spoke about this phenomenon with Dr Genevera Allen from Rice University in Houston, who believes that machine learning may be to blame for exacerbating science's reproducibility crisis. "Often these studies are not found out to be inaccurate until there's another real big dataset that someone applies these techniques to and says 'oh my goodness, the results of these two studies don't overlap'," she explained.