ARIA has this example:

<div role="img" aria-labelledby="caption">
  <img src="example.png" role="presentation" alt="">
  <p id="caption">A visible text caption labeling the image.</p>
</div>

This can literally be transformed into a <figure>:

<figure role="img" aria-labelledby="caption">
  <img src="example.png" role="presentation" alt="">
  <figureCaptionElement id="caption">
   A visible text caption labeling the image.
  </figureCaptionElement>
</figure>

However, for some reason Validator.nu does not permit this, despite that HTML5 does *not* define socalled "strong native semantics" for <figure>