🤖 machine learning

Position: AI for Science Should Treat Measurement-to-Dataset Pipelines as Inference Components

This paper argues that AI for Science workflows must treat measurement-to-dataset pipelines as active inference components rather than fixed data sources to address hidden hypothesis spaces, uncertified transportability, and ungoverned multiplicity, thereby enabling quantifiable pipeline adequacy and reproducible scientific evidence.

Original authors: Ling Zhan, Xiaoyao Yu, Tao Jia

Published 2026-05-26

📖 6 min read🧠 Deep dive

CC BY 4.0

Original authors: Ling Zhan, Xiaoyao Yu, Tao Jia

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

The Big Idea: The "Frozen Lens" Problem

Imagine you are trying to understand a mysterious object in a dark room. You can't see the object directly; you can only see its shadow cast on the wall by a specific lamp.

In AI for Science, researchers often act like they are studying the object itself. But in reality, they are studying the shadow.

The paper argues that in many scientific fields (like brain imaging, astronomy, or drug discovery), the data scientists use isn't "raw reality." It is a processed, cleaned-up, reconstructed version of reality. This processing happens through a long chain of steps called a pipeline (filtering noise, filling in missing pieces, converting signals).

Currently, scientists treat the final dataset as if it were a fixed, objective fact. They say, "Here is the data, now let's build a model." The authors call this the "Frozen Lens." They argue this is a mistake. The pipeline that created the data is actually a huge part of the scientific reasoning, and by "freezing" it (ignoring it), we are hiding the fact that our conclusions depend entirely on how we chose to cast that shadow.

The Three Ways This Goes Wrong

The authors identify three specific ways this "Frozen Lens" causes science to fail. Think of these as three traps:

1. The Hidden Menu (Hidden Hypothesis Space)

The Analogy: Imagine a restaurant that serves you a burger. You assume the burger is the "truth" about what the chef made. But the chef had a menu of 50 different ways to make that burger (different buns, different cooking times, different sauces). The restaurant only served you one specific version and didn't tell you about the other 49.
The Problem: When scientists release a dataset, they usually pick one way to process the raw measurements. They don't tell you that if they had chosen a slightly different filter or a different cleaning method, the "burger" (the data) might have looked completely different.
The Result: The AI model learns to predict based on that one specific version of the burger. It thinks it has discovered a universal truth, but it's actually just memorizing the quirks of that one specific recipe. The "menu" of other possible realities is hidden.

2. The Broken Map (Uncertified Transportability)

The Analogy: Imagine you have a map of a city that was drawn perfectly for a sunny day. You use this map to navigate. But then, it starts raining, and the streets flood. The map doesn't show you the puddles because it was only drawn for sunny days. You try to use the map in the rain, get stuck, and blame your driving skills.
The Problem: A pipeline might work great in the lab where the data was collected (the "sunny day"). But if you try to use that same data processing method on data from a different hospital, a different country, or a different machine (the "rainy day"), the pipeline might break or distort the data in weird ways.
The Result: Scientists often don't know when their data processing method stops working. When an AI model fails on new data, we don't know if the AI is bad, or if the "map" (the pipeline) was just invalid for that new environment.

3. The Crowd of Experts (Ungoverned Multiplicity)

The Analogy: Imagine a jury of 100 experts looking at the same crime scene photo. They all agree on the facts, but they all use slightly different magnifying glasses. 90 of them conclude the suspect is guilty. 10 conclude the suspect is innocent. The news reports, "The experts say the suspect is guilty," ignoring the 10 who disagreed.
The Problem: In science, there are often many "defensible" ways to process data. You can choose to filter out noise in five different valid ways. If you pick one, you get one result. If you pick another, you get a different result.
The Result: Scientists often pick the one method that gives them the "coolest" or most significant result and ignore the fact that 99 other valid methods would have given a different answer. They treat their single choice as the absolute truth, rather than admitting, "We tried 100 ways, and the answer is a bit shaky."

The "EEG Audit": A Reality Check

To prove this isn't just theory, the authors ran a massive experiment on brain data (EEG) related to depression.

The Setup: They took raw brain wave data and ran it through 245,376 different combinations of processing pipelines. (Imagine trying every possible combination of filters, cleaning tools, and measurement rules).
The Goal: They were looking for a specific signal: a difference in brain activity between people with depression and healthy people.
The Shocking Result:
- Many of the 245,000+ pipelines found a "significant" difference.
- BUT, when they checked if those differences held up across different groups of people (different datasets), only ONE pipeline out of 245,376 survived.
- That is a survival rate of 0.0004%.

What this means: Almost every time scientists found a "discovery" in this field, it was likely an accident of which specific pipeline they happened to pick. The "truth" was so fragile that changing the processing rules slightly made the discovery disappear.

The Solution: The "Computable Observation Framework"

The authors aren't saying we should stop doing science. They are saying we need to change how we treat data.

Current Way:

Raw Data $\rightarrow$ [Black Box Pipeline] $\rightarrow$ "Clean Data" $\rightarrow$ AI Model.
We pretend the "Clean Data" is the truth.

Proposed Way:

Raw Data $\rightarrow$ The Pipeline as a Variable $\rightarrow$ AI Model.
We treat the pipeline like a scientific hypothesis. We ask: "Does this conclusion hold up if we change the pipeline?"

They propose building a system where:

Pipelines are "Executable Objects": Instead of just a PDF report saying "we used Method A," the code for Method A (and all its possible variations) is saved as a digital object that can be run, tested, and changed.
Stability is the Goal: Instead of just looking for the model that predicts best, we look for the model that gives the same answer no matter which valid pipeline we use.
Uncertainty is Shared: If 100 pipelines give 100 different answers, we report that uncertainty. We don't hide it.

Summary

The paper argues that in AI for Science, the tool used to measure reality is just as important as the measurement itself. By ignoring the tool (the pipeline), scientists are often building AI models on shaky ground. They need to stop treating processed data as a fixed fact and start treating it as a variable part of the scientific experiment.

Technical Summary: AI for Science Should Treat Measurement-to-Dataset Pipelines as Inference Components

Problem Statement
In "AI for Science" (AI4Science) workflows, particularly those relying on indirect observation (e.g., neuroimaging, structural biology, radio astronomy), the scientific target is inferred from raw sensor measurements through multi-stage measurement, reconstruction, and preprocessing pipelines. Currently, these workflows treat the released dataset as a fixed interface, effectively "freezing" the observation model ( $P_\phi$ ) that maps raw measurements ( $x$ ) to a derivative representation ( $z$ ).

The authors argue that treating these pipelines as static data definitions rather than active inference components creates a "frozen lens." This approach obscures the uncertainty inherent in pipeline choices and leads to three specific failure modes:

Hidden Hypothesis Space (C1): The released dataset does not specify the pipeline configuration or its validity conditions. Consequently, the effective hypothesis space is restricted to explanations compatible only with the frozen representation, hiding alternative observation models that could yield different mechanistic conclusions.
Uncertified Transportability (C2): Documented pipelines rarely come with a tested regime of validity. When acquisition protocols, sensors, or noise conditions shift, the pipeline may fail to transport the representation correctly. Because preprocessing is often lossy and many-to-one, downstream learners cannot recover distinctions removed upstream, leading to brittle mechanistic claims that do not generalize across regimes.
Ungoverned Multiplicity (C3): Multiple defensible pipelines often exist for a given task. Current practices rarely propagate the dispersion among these pipelines into uncertainty-qualified evidence. Instead, a single "default" or "best-practice" pipeline is selected, collapsing valid variation into a point estimate and inflating confidence in mechanistic claims.

Methodology
To empirically validate these claims, the authors conducted a large-scale audit in the domain of EEG functional connectomics, specifically investigating hemispheric asymmetry as a biomarker for Major Depressive Disorder (MDD).

Experimental Design: The audit evaluated a vast space of admissible observation models. They combined 864 distinct preprocessing pipelines (varying in filtering, referencing, denoising, downsampling, and windowing) with 284 Statistical Pairwise Interaction (SPI) operators (rules for converting channel time series into functional connectivity edge weights).
Scale: This resulted in 245,376 unique pipeline-operator configurations tested across two independent datasets (MODMA and CISIR).
Evaluation Criteria:
- Significance: Within-dataset separation between MDD and Healthy Control (HC) groups using a Mann–Whitney U test.
- Stability: Cross-dataset agreement in effect direction and relative ordering, assessed via bootstrap resampling.
Framework Proposal: The paper proposes a shift from procedural checklists to Computable Observation Frameworks (COFs). This framework consists of three layers:
1. Representation: Versioned, executable specifications of the admissible pipeline family ( $\Phi$ ), including operator graphs and parameter domains.
2. Optimization: Search procedures that optimize for stability and transportability across settings rather than just predictive fit on a frozen artifact.
3. Governance: Platforms for reusable audit records that accumulate evidence over the observation-model space, allowing for the propagation of uncertainty.

Key Results
The empirical audit yielded a stark survival rate for stable mechanistic evidence:

While many configurations achieved statistical significance within a single dataset, the vast majority failed to maintain consistent conclusions across the two independent datasets.
Only one configuration out of 245,376 satisfied both the within-dataset significance and cross-dataset stability criteria.
This corresponds to a survival rate of approximately 0.0004%.
The results demonstrate that the "frozen lens" creates a landscape where significant findings are highly sensitive to pipeline choices, and standard selection heuristics often pick pipelines that do not generalize, thereby validating the existence of C1, C2, and C3.

Significance and Claims
The paper claims that the current paradigm of treating datasets as fixed interfaces is epistemically insufficient for AI4Science, where the goal is mechanistic discovery rather than mere prediction.

Reframing Pipelines: The authors assert that pipelines are not merely preprocessing steps but are inference components. Their outputs should be treated as uncertain evidence rather than objective facts.
Addressing the "Frozen Lens": By making pipelines computable objects, the community can quantify pipeline adequacy and stability. This converts implicit implementation choices into auditable, reproducible, and cumulative scientific evidence.
Role of ML: The paper posits that machine learning's role should extend beyond optimizing predictive models ( $\theta$ ) to governing the observation model ( $\phi$ ). This involves representing design spaces, defining stability-aligned objectives, and performing budgeted search to select or aggregate across admissible pipelines while propagating residual uncertainty.
Call to Action: The authors call for the development of domain-specific Computable Observation Frameworks to replace static "best practice" checklists with dynamic, queryable layers of the inference workflow. This shift is presented as necessary to prevent the accumulation of brittle, non-replicable scientific claims as AI4Science scales via foundation models and automated laboratories.