The latent vector field can be used to analyze the training dynamics of neural networks, giving insights on how representations form during the training process.

Attractors can be recovered from noise, enabling us to probe the knowledge encoded in the weights of vision foundation models, without requiring any input data.

Inductive biases in the training process cause the formations of attractors in the latent vector field, characterizing memorization and generalization regimes of the network.

Neural networks implicitly define a latent vector field on the data manifold, via autoencoding iterations🌀

This representation retains properties of the model, revealing memorization and generalization regimes, and characterizing distribution shifts

📜: arxiv.org/abs/2505.22785

Attractors can be recovered from noise, enabling us to probe the knowledge encoded in the weights of vision foundation models, without requiring any input data.

Inductive biases in the training process cause the formations of attractors in the latent vector field, characterizing memorization and generalization regimes of the network.

LFMs enable to stitch together neural network modules, allowing sample efficient information transfer with minimal to no prior correspondence. Additionally, LFMs can be applied on top of any off-the-shelf parametric transformation method.

(4/N)

LFMs find correspondences from very few prior knowledge: nearly isometric spaces can be aligned reducing the number of known correspondences by a factor of 100 (from 300 to 3 in the example below).

(3/N)

LFMs allows to (i) robustly compare latent representations, and (ii) to identify local distortions induced by the map, providing an interpretable measure of representational similarity.

(2/N)

Excited to present "Latent Functional Maps" at #NeurIPS !

We show how neural models can be aligned by matching function spaces on representation manifolds, providing a unified framework for model comparison, matching, and information transfer.

📜: arxiv.org/abs/2406.14183

👇🧵