Jethro's Braindump

Multi-modal Alignment

Multi-modal alignment is a sub-challenge in Multi-modal Machine Learning, involving the finding of relationships and correspondences between sub-components of instances from two-or-more modalities. For example, we want to find areas of the image corresponding to the caption’s words or phrases.

In explicit multi-modal alignment, the objective of aligning sub-components in modalities is of interest. In implicit multi-modal alignment, this is typically an intermediate step for another task.

Explicit Alignmnt

Explicit alignment algorithms can be further split into unsupervised and weakly-supervised algorithms.

The unsupervised approaches do not require or use any direct alignment labels. Dynamic Time Warping is a dynamic progamming approach to align multi-view time series.

Implicit Alignment

Implicit alignment allows for better performance in downstream tasks. Both graphical models and neural network approaches are used here. Examples of tasks where implicit alignment helps include machine translation.

Links to this note