Latent Space Translation via Semantic Alignment
This work addresses the challenge of aligning latent representations across diverse models, which is incremental but useful for researchers and practitioners in machine learning seeking efficient model integration.
The paper tackles the problem of translating latent spaces between different pre-trained neural networks, showing that simpler transformations than previously thought can effectively stitch encoders and decoders without additional training, enabling zero-shot multimodal tasks like text-vision classification.
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.