How to use AI for discovery without leading science astray

In the last decade, artificial intelligence (AI) has played a significant role in scientific research. Machine learning models have been employed to predict protein structures, estimate deforestation levels in the Amazon rainforest, and classify distant galaxies to identify potential exoplanets. However, although AI has the potential to accelerate scientific discovery, there is a risk of misleading or false results. Machine learning models, like chatbots that sometimes produce fictitious responses, can sometimes produce inaccurate outcomes.

To address this issue, researchers at the University of California, Berkeley, have introduced a new statistical technique called prediction-powered inference (PPI). This technique allows scientists to use predictions obtained from machine learning models while correcting for potential errors and biases. In this article, we will explore how PPI works, its applications in various scientific domains, and its significance in data-intensive research.

When conducting scientific experiments, researchers aim to obtain a range of plausible answers instead of a single definitive answer. They achieve this by calculating confidence intervals, which assess the variability of results obtained through repeated experiments. However, machine learning systems focus primarily on individual data points and cannot provide scientists with the uncertainty assessments they require.

For instance, AlphaFold, a popular machine learning model used for predicting protein structures, can only offer a single structure prediction without any measure of confidence. Scientists may be tempted to treat these predictions as data and compute classical confidence intervals, disregarding the fact that machine learning models have hidden biases that come from the training data. Such biases can skew the results, especially when exploring phenomena at the boundaries between known and unknown realms.

To address these limitations, researchers at UC Berkeley developed a technique called Prediction-Powered Inference (PPI). PPI leverages a small amount of unbiased real-world data to correct the output of large, general models like AlphaFold, specifically in the context of scientific inquiries. By combining these two sources of evidence, PPI enables the formation of valid confidence intervals.

The central idea behind PPI is to combine the predictions from machine learning models with unbiased data related to the specific hypothesis being investigated. This approach allows scientists to leverage the benefits of machine learning models while correcting for potential errors and biases. The key advantage of PPI is its ability to provide reliable confidence intervals even when the nature of errors in the machine learning model is unknown at the outset.
Applying PPI in Scientific Research

PPI has proven to be effective in various scientific domains, ranging from environmental studies to astrophysics and genetics. Let's delve into some notable applications of PPI in these fields:

• Environmental Studies: Estimating Deforestation Levels in the Amazon

One of the uses of PPI is in estimating deforestation levels in the Amazon rainforest by utilizing satellite imagery. Machine learning models trained on satellite data can accurately identify deforestation in specific regions of the forest. However, when combined to estimate deforestation across the entire Amazon, these models can yield skewed confidence intervals due to their inability to recognize newer patterns of deforestation. PPI helps to correct this bias by incorporating a small number of human-labeled regions of deforestation.

• Protein Folding: Predicting Protein Structures with Confidence

Protein folding is a critical process in comprehending protein function and designing therapeutics. AlphaFold, a machine learning model, has shown remarkable success in predicting protein structures. However, it cannot provide confidence intervals for its predictions. By applying PPI, scientists can integrate additional unbiased data to obtain valid confidence intervals for protein structures predicted by AlphaFold.

• Astrophysics: Classifying Distant Galaxies

Machine learning models have been utilized to classify distant galaxies based on their characteristics, aiding in the search for exoplanets. However, these models might generate unrealistic output when faced with phenomena at the edge of our current knowledge. PPI offers a solution by allowing scientists to correct potential errors and biases in the models using unbiased data.

• Genetics: Exploring Gene Expression Levels

Understanding gene expression levels is crucial in deciphering genetic mechanisms and their role in various biological processes. Machine learning models can be used to predict gene expression levels based on diverse factors. PPI enables scientists to incorporate additional unbiased data to obtain valid confidence intervals for gene expression predictions.

Other Applications of PPI

The potential applications of PPI are vast and not limited to the examples mentioned above. The technique can be utilized in diverse research fields, including but not limited to plankton counting, investigating the relationship between income and private health insurance, and exploring other scientific phenomena.

AI has transformed scientific research by speeding up processes and enabling predictions in various domains. However, it is crucial to use AI models with caution, considering the potential for misleading or false results. UC Berkeley researchers have developed the prediction-powered inference (PPI) technique to overcome this challenge, enabling scientists to use machine learning models while correcting for errors and biases.

PPI incorporates a small amount of unbiased data to form valid confidence intervals, providing scientists with the necessary uncertainty assessments. By applying PPI, scientists can leverage the power of AI models in diverse scientific inquiries, ranging from environmental studies to astrophysics and genetics. This technique contributes to the advancement of modern data-intensive, model-intensive, and collaborative science.

With the advent of PPI, researchers can confidently embrace AI for scientific discovery, knowing that they can mitigate potential errors and biases. The future of AI in scientific research appears promising, with PPI becoming an integral component of data analysis, hypothesis testing, and decision-making processes. As scientists continue to unlock the mysteries of the natural world, PPI will play a crucial role in ensuring the reliability and validity of AI-driven insights and predictions.