Last week, Zac Spalding (@zspald.bsky.social‬, 4th year PhD student, @dukeubme.bsky.social‬) presented Adam Gosztolai, Robert Peach, and colleagues’ 2025 paper on MARBLE, a method for finding interpretable latent representations of neural dynamics.

🤍1️⃣: The initial proximity graph is a clever way to define distance and neighborhoods between inputs that can be used for downstream training. 🤍2️⃣: The rotation invariance is important and likely useful for extracting shared latent representations from systems with minor differences.

🤍3️⃣: The comparisons to state-of-the-art latent dynamical systems models are great for properly contextualizing the performance of MARBLE.

❔1️⃣: It is stated that non-neighbors (both within and across manifolds) are negative samples (mapped far) during the contrastive learning step. Does treating non-neighbors within and across manifolds as similarly “distant” lead to less interpretability of larger distances in latent space?

❔2️⃣: It seems that a linear transformation between MARBLE representations of different animals was necessary because the same information is present in the latent space but not necessarily with the same ordering... (con't)

If separate animals were treated as separate manifolds with an embedding-agnostic MARBLE, would you still expect an informative latent space to be learned without any need for post-hoc alignment?