Create and explore imaginative 3D worlds with our new product Dreamfield (opens in a new tab)

Criteria for Selecting Embeddings to Drive Active Learning

Establish principled criteria for selecting an embedding representation to drive active learning in biodiversity monitoring, given that downstream performance is sensitive to the choice of embedding.

Background

The review describes a growing use of pretrained encoders and embeddings as the basis for active learning in biodiversity monitoring. In this setting, active learning selects an informative subset of an unlabelled pool for annotation, while the encoder may be frozen or jointly fine-tuned with a classification head.

Because downstream performance depends on the embedding representation, choosing an embedding is itself an important methodological decision. The review states that principled criteria for making this choice have not yet been established, identifying the issue as an open question in the context of adapting pretrained models for biodiversity monitoring.

References

Downstream performance is sensitive to the choice of embedding \citep{dumoulin2025search}, yet principled criteria for selecting an embedding to drive AL remain to be established; we identify this as an open question.

— Active Learning for Biodiversity Monitoring: From Label Efficiency to Reliable Ecological Inference  (2609.27409 - McEwen et al., 23 Sep 2026) in Section 3.2, subsection “Models”

When FINE is applied to this setting, the choice of representation remains an open question.

— Learning What to Trust in Multimodal Learning under Noisy Supervision  (2610.11057 - Zou et al., 8 Oct 2026) in Section 3.1, subsection “FINE for Learning with Noisy Labels: A Recap”