[Paper Review] Is machine learning good or bad for the natural sciences?

David W. Hogg, Soledad Villar|arXiv (Cornell University)|May 28, 2024

Big Data and Business Intelligence7 citations

TL;DR

The paper argues that ML has both valuable roles and potential pitfalls in natural sciences, detailing two major biases it can introduce and proposing safe, causal-aware usage patterns.

ABSTRACT

Machine learning (ML) methods are having a huge impact across all of the sciences. However, ML has a strong ontology - in which only the data exist - and a strong epistemology - in which a model is considered good if it performs well on held-out training data. These philosophies are in strong conflict with both standard practices and key philosophies in the natural sciences. Here we identify some locations for ML in the natural sciences at which the ontology and epistemology are valuable. For example, when an expressive machine learning model is used in a causal inference to represent the effects of confounders, such as foregrounds, backgrounds, or instrument calibration parameters, the model capacity and loose philosophy of ML can make the results more trustworthy. We also show that there are contexts in which the introduction of ML introduces strong, unwanted statistical biases. For one, when ML models are used to emulate physical (or first-principles) simulations, they amplify confirmation biases. For another, when expressive regressions are used to label datasets, those labels cannot be used in downstream joint or ensemble analyses without taking on uncontrolled biases. The question in the title is being asked of all of the natural sciences; that is, we are calling on the scientific communities to take a step back and consider the role and value of ML in their fields; the (partial) answers we give here come from the particular perspective of physics.

Motivation & Objective

• Describe the fundamental ontology and epistemology of machine learning and contrast it with natural sciences.
• Identify two strong statistical biases ML can introduce into natural-science research.
• Map safe contexts where ML enhances scientific practice and argue for cautious, causal-aware usage.
• Encourage natural-science communities to evaluate the role of ML and adopt practices that preserve scientific understanding.

Proposed method

• Define a broad ML ontology (data-centric) and contrast with the latent-structure focus of natural sciences.
• Explain how ML’s epistemology centers on held-out data performance rather than latent interpretability.
• Identify and explain two biases: emulator-induced confirmation bias and training-set bias amplification.
• Provide examples of safe ML applications in real-time decisions, nuisance modeling, and causal inference.
• Discuss contexts where ML can be beneficial (e.g., foregrounds, calibrations, rare-object discovery) and where it can be harmful.

Experimental results

Research questions

• RQ1 What roles can ML play in advancing natural-science understanding and discovery?
• RQ2 What are the main biases ML introduces in natural-science analyses, and can they be mitigated?
• RQ3 In which contexts can ML provide safe, beneficial contributions without compromising understanding?
• RQ4 How should natural-science communities adopt ML tools to preserve epistemic standards?

Key findings

• ML has valuable places in contemporary science, particularly in operational and causal-context uses.
• Two major biases are introduced by ML in natural sciences: emulator-induced confirmation bias and training-set bias amplification.
• These biases can be hard to correct and often arise when using ML-generated labels or emulators in downstream analyses.
• In causal settings, expressive ML models can yield more conservative and robust conclusions about causation when modeling confounders.
• There are numerous safe and even required uses of ML in natural sciences, including real-time decisions, nuisance modeling, and detection of outliers, provided epistemic standards are maintained.