What does it mean to model an embedding as a distribution?

Instead of mapping an input to a single vector, the encoder outputs the parameters of a probability distribution, typically a Gaussian with a mean and a variance. The mean plays the role the point embedding used to play, and the variance expresses how much the encoder commits to that location. Specific, unambiguous content yields a tight distribution; broad or polysemous content yields a wide one. Similarity is then computed between distributions, for example with a KL based objective, rather than as a single distance between points.

Why does calibrated uncertainty matter for retrieval?

A conventional ranker always returns a ranked list and gives no principled signal about whether the top result is genuinely a good match or merely the least bad option available. When embeddings carry variance, the margin between distributions provides a confidence estimate alongside the ranking, so a downstream system can abstain, ask a clarifying question, or fall back to another strategy when the match is weak. For an agent acting on retrieval results without a human reviewing them, knowing when the system does not know is often worth more than a small gain in recall.

Your Embedding Is a Point. It Should Be a Cloud.

Name: Your Embedding Is a Point. It Should Be a Cloud.
Uploaded: 2026-08-01T20:00:00Z
Duration: 60 s

Multimodal Search in Action

Mixpeek Team

August 1, 2026

Summary

Every retrieval system treats an embedding as a point: one vector, one location, one distance to the query. That assumption hides something, because content is not equally specific. A five second clip of someone walking is unambiguous; a two minute clip of a wedding is many things at once. Both collapse to a single point, and the system ends up exactly as confident about each. The Distribution-Alignment Bridge (Chae, Lee, Ahn, ECCV 2026) models text and video embeddings as Gaussians defined by a mean and a variance, so a vague caption becomes a wide distribution and a precise one becomes tight. Matching becomes distribution alignment rather than point distance, trained with a KL based distribution aware contrastive loss and a deterministic diffusion style bridge that refines the text distribution toward the video distribution over a truncated number of steps, evaluated on MSR-VTT, MSVD and VATEX. The payoff is calibrated uncertainty in the ranking. Paper: arxiv.org/abs/2607.20984

text-to-video-retrievalembeddingsuncertaintyeccvvideo-searchcomputer-vision

Your Embedding Is a Point. It Should Be a Cloud.

Summary

About this video

Frequently Asked Questions