A survey detection channel overrides the pixels in an astronomical foundation model, and biases tomographic mean redshifts
AION-1, a 39-modality astronomical foundation model trained on 200M+ objects, inherits systematic biases from incomplete survey catalogues used during training Editing only the survey segmentation map (while keeping image tokens identical) changes all model outputs—flux, size, ellipticity, redshift—by 110-4400× compared to a matched placebo The bias mechanism is "detection gating": the model responds to whether a detection exists at the field centre (r = 0.47) rather than the actual light within
Analysis
TL;DR
- AION-1, a 39-modality astronomical foundation model trained on 200M+ objects, inherits systematic biases from incomplete survey catalogues used during training
- Editing only the survey segmentation map (while keeping image tokens identical) changes all model outputs—flux, size, ellipticity, redshift—by 110-4400× compared to a matched placebo
- The bias mechanism is "detection gating": the model responds to whether a detection exists at the field centre (r = 0.47) rather than the actual light within the mask (r = 0.30), and ignores how pipelines partition blended sources (R = -0.006)
- Contradicted catalogue photometry degrades model performance nine times worse than providing no metadata at all
- The Legacy Survey pipeline leaves 3.68% of targets with no segmentation covering their position, which propagates to shift tomographic mean redshifts by a median 0.71× the LSST DESC requirement and exceeds it in 12 out of 40 assignments
Why It Matters
This work reveals a critical vulnerability in astronomical foundation models: they can amplify and inherit systematic errors from the survey pipelines that produced their training data, potentially biasing cosmological measurements at levels that exceed mission requirements. For AI practitioners building multimodal models on observational data, it demonstrates that incomplete or contradictory metadata channels can dominate model behavior in ways that are difficult to detect without causal intervention audits.
Technical Details
- Model audited: AION-1, a 39-modality transformer trained on over 200 million astronomical objects, combining image tokens with catalogue-derived metadata across multiple modalities
- Causal intervention methodology: The authors held image tokens byte-identical while editing only the survey segmentation map, measuring the resulting output changes against a matched placebo—this isolates the causal effect of the detection channel from the raw pixel data
- Detection gating mechanism: The model's predictions are driven primarily by the presence/absence of a detection at the field centre (correlation r = 0.47) rather than the photometric content within the mask (r = 0.30); for real blended sources, the model is effectively blind to pipeline partitioning decisions (R = -0.006)
- Tokeniser limitations: The image codec resolves only 28 effective states on source patches versus 934 for the spectrum codec, and redshift readout is quantisation-limited, creating an information bottleneck
- Mitigation findings: Withholding the detection channel removes the bias at no measurable performance cost, spectroscopy eliminates the effect entirely, and the bias grows with model scale—suggesting larger models are more, not less, vulnerable to catalogue systematics
Industry Insight
- Foundation models trained on pipeline-derived products risk inheriting and amplifying systematic errors; practitioners should audit metadata channels with causal interventions, not just correlation analysis, before deploying models for precision science
- The finding that contradicted metadata is worse than no metadata at all suggests a general principle for multimodal training: when modalities conflict, the model may latch onto the most confidently encoded signal rather than reconciling them—consider training with explicit conflict resolution or dropping unreliable channels
- As astronomical foundation models scale up, bias from incomplete catalogues will intensify; investing in spectroscopic validation and designing models that can operate without detection-channel metadata should be prioritized for next-generation surveys like LSST
Disclaimer: The above content is generated by AI and is for reference only.