Reproduce “Efficient Quantization-Aware Distillation with Cross-Modal Alignment for Edge Vision-Language Models” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Jinwoo Jeon, GyuYeop Do, Yubin Lim, Nam-Joon Kim, Hyun Gon Ryu, Hyuk-Jae Lee, Byung-Jun Lee
An efficient training framework combines quantization-aware distillation with cross-modal attention to adapt CLIP-style vision–language models for edge devices. By anchoring low-precision student embeddings to frozen teacher representations and transferring RGB semantics to non-RGB inputs, the method improves non-RGB and multimodal recognition while reducing training time by 81% compared with EdgeVL.
Key insights
- Anchoring quantized student features directly to frozen teacher embeddings helps preserve the geometry and semantics of the CLIP representation space; pseudo-label supervision further prevents same-class examples from being treated only as negatives.
- Cross-modal attention improves non-RGB representations more flexibly than simple feature averaging while adding only a single attention layer, though it increases memory use and reduces throughput.
- The benefits are architecture- and dataset-dependent: the method generally improves in-domain performance and several cross-dataset results, but Swin-T shows inconsistent transfer gains on unseen datasets.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.