Michelle has been invaluable in that role for nearly a decade now, but the time has come for @aoir.bsky.social to find a new coordinator to take on these vital duties. 2/2

Job posting is on our listserv. Please share & apply! listserv.aoir.org/pipermail/ai...

[…] symbolically aligning their influence and teamwork with the collective efforts needed to address environmental challenges[…]" even without any cleaning the text gives us a clear semantically accurate description of the meaning of the picture that would not be possible otherwise.

"[…] a powerful visual blend of modern celebrity culture with pressing environmental issues, using the popular music group as a metaphorical representation of various solutions to climate change. Each member is labeled with a different aspect of sustainability

fandom dynamics and read the text that goes with it. The actual meaning ("these are the solutions we need for the energy transition") require some cultural work and is hardly connected with the surface features (not even with the text). Gpt4o offered this (shortened) description:

awareness and why it matters.
Take a look at this picture (that we got from a dataset of climate-related activism). This photo is remarkably complex from a cultural point of view, yet very simple in its message. To understand it you need to know who the BTS are, to get an intuition of k-pop

different configurations can affect the results and we are now quite confident that this could be a valuable strategy to consider if you want to computationally study visual data in the context of social sciences.

(e.g. recognising celebrities, symbols etc); 2) having an intermediate textual step allows for an easier interpretability of the clusters that do not need to be visually inspected.

The whole pipeline it's rather complex and contains many moving parts, we've conducted extensive testing of how

clustering images that have the "same meaning" rather than showing the same "surface-level visual features". We show that our proposed approach achieves a considerably higher clustering quality, as well as offering additional positive aspects: 1) VLLM show high level of cultural awareness

Here we present and evaluate a strategy that uses a first step of image description done via VLLM and then clusters the resulting texts via text-embeddings. We also compare this strategy against a traditional CNN-based approach given the task of "semantic clustering" that we define as