A team led by Princeton undergraduate Veena Krishnaraj has demonstrated that neural networks trained via transfer learning can absorb the standard ΛCDM model of cosmology using cheap simulations, then reduce the number of expensive beyond-standard-model runs by more than tenfold. The results, published in the Journal of Cosmology and Astroparticle Physics, are a clear win for computational efficiency—especially with upcoming data deluges from the Vera C. Rubin Observatory and Euclid. But the work also exposes a critical failure mode: the networks carry forward biases from pre-training, performing well when new physics is far from standard parameter space but stumbling when signals overlap with degenerate directions already mapped inside ΛCDM.

Degeneracies—where different parameter combinations produce nearly identical observable effects on the cosmic web—are a known headache in cosmology. Neutrino mass and matter fluctuation amplitude trade off in ways that look identical in large-scale structure data. Pre-trained networks amplified those trade-offs, treating slight deviations as noise rather than evidence of new physics. The team calls this “negative transfer,” and it’s not random. It’s driven by physical degeneracies baked into the model. The AI became overconfident in regions it had already mastered, down-weighting genuine signals that sat inside those degenerate zones.

From a systems design perspective, this is a classic overfitting-to-priors problem. Transfer learning slashed compute costs—a major win for teams running limited simulation budgets—but the trade-off is that the model’s latent representations encode the standard model’s blind spots. The authors suggest mitigation strategies: hybrid training schedules that inject adversarial noise, or Bayesian neural networks that propagate uncertainty estimates forward, flagging degenerate parameter spaces for human review or additional simulation runs.

Francisco Villaescusa-Navarro of the Flatiron Institute frames the stakes clearly: transfer learning can accelerate discovery, but without deliberate safeguards, it may steer the field away from the anomalies it was built to find. Some degeneracies are physical; others are baked into network weights. Distinguishing the two will require new techniques—perhaps networks that actively unlearn or ensembles trained on contradictory priors.

For data and ML engineers working in scientific domains, the lesson is straightforward: pre-training on standard models is a powerful accelerator, but it can also create a cage. The network’s inherited assumptions must be audited, stress-tested, and sometimes deliberately broken. The Princeton team has given the community a map of the trapdoors. How we design around them will determine whether AI becomes a genuine discovery partner or a faster way to confirm what we already believe.

Source: Webpronews