Hi Ben. Yes indeed, they require a certain number of observations. However, crucially, it is not trivial to "influence" what criteria are used perform the clustering (e.g. in the case of images, the visual features). Our approach enables the discovery of interpretable features that can be clustered.

There is more info in our new preprint:

INQUIRE-Search: A Framework for Interactive Discovery in Large-Scale Biodiversity Databases

arxiv.org/abs/2511.15656

The full paper can be found here:

Interpretable Text-Guided Image Clustering via Iterative Search
arxiv.org/abs/2506.12514

Today at #BMVC2025, Bingchen Zhao will present our new text-guided clustering approach uses an iterative discovery process, guided by an unsupervised clustering objective, to generate interpretable visual concepts that better capture the criteria expressed in a user's instructions.

This project is in collaboration with Rowland Kao @rowlandkao.bsky.social

The PhD Programme has some information sessions coming up (including one today).
21st November, 14:30-15:30 GMT
9th December, 10:00-11:00 GMT
lnkd.in/egccTN2d

Finally, there is also the European Lab for Learning and Intelligent Systems (ELLIS) programme.

ELLIS PhD Programme
ellis.eu/news/ellis-p...
Deadline 31st Oct 2025

In addition there is

CDT in Dependable and Deployable AI for Robotics
www.cdt-d2air.uk
Deadline 14th Nov 2025

CDT in Designing Responsible Natural Language Processing
www.responsiblenlp.org
Deadline 7th Jan 2026

We are fortunate to have several Centres for Doctoral Training (CDTs) here which cover full funding and scholarships for 4 years, e.g.,

CDT in Machine Learning Systems
informatics.ed.ac.uk/cdt-in-machi...
Deadline 11th Dec 2025

We have some fantastic invited speakers lined up:
sites.google.com/g.harvard.ed...

In addition, we have a paper track for both novel unpublished and published work (deadline Oct 10th):
sites.google.com/g.harvard.ed...

Very cool!

Check out the paper for more details:

Feedforward Few-shot Species Range Estimation
arxiv.org/abs/2502.14977
Lange et al. ICML 2025

This work was led by Christian Lange, with support from our amazing collaborators Max Hamilton, Elijah Cole, Alexander Shepard, Samuel Heinrich, Angela Zhu, Subhransu Maji, Grant Van Horn, and Oisin Mac Aodha.

FS-SINR is efficient. At test time, it can take an arbitrary number of observations (i.e., context locations) as input, along with optional metadata, and generate a predicted range in a single forward pass of the model.