Optimization Encoders: Rethinking Second-Order Meta-Learning for Neural Fields
The paper frames latent gradient adaptation itself as the encoder for conditional neural fields.
It argues that second-order meta-learning exposes a learning pathway that first-order approximations drop. The authors introduce MetaLF, an equivariant transformer neural field that uses self-attention over a latent pointcloud to coordinate local observations into non-local structure. In experiments, MetaLF improves image and 3D shape reconstruction within three to five gradient updates and supports semantic prediction across images, shapes, and volumes. ArXiv · AI/CL/LG's note
It argues that second-order meta-learning exposes a learning pathway that first-order approximations drop. The authors introduce MetaLF, an equivariant transformer neural field that uses self-attention over a latent pointcloud to coordinate local observations into non-local structure. In experiments, MetaLF improves image and 3D shape reconstruction within three to five gradient updates and supports semantic prediction across images, shapes, and volumes. ArXiv · AI/CL/LG's note
score 4