Latent Space Translation via Semantic Alignment
Keywords: latent space translation, relative representation, Procrustes analysis, zero-shot, stitching, latent communication, representation learning, manifold alignment, multimodal
TL;DR: A semantic correspondence between latent spaces is enough to unlock translation between them, reaching zero-shot stitching of arbitrarly pre-trained neural components across various tasks, modalities and architectures.
Abstract: While different neural models often exhibit latent spaces that are alike when exposed to semantically related data, this intrinsic similarity is not always immediately discernible. Towards a better understanding of this phenomenon, our work shows how representations learned from these neural modules can be translated between different pre-trained networks via simpler transformations than previously thought. An advantage of this approach is the ability to estimate these transformations using standard, well-understood algebraic procedures that have closed-form solutions. Our method directly estimates a transformation between two given latent spaces, thereby enabling effective stitching of encoders and decoders without additional training. We extensively validate the adaptability of this translation procedure in different experimental settings: across various trainings, domains, architectures (e.g., ResNet, CNN, ViT), and in multiple downstream tasks (classification, reconstruction). Notably, we show how it is possible to zero-shot stitch text encoders and vision decoders, or vice-versa, yielding surprisingly good classification performance in this multimodal setting.
Supplementary Material: zip
Submission Number: 12209