Huawei, Sirius, Semantic extraction and synthesis
Our task was to investigate methods of dimension reduction methods and show that for the presented data there exist such embedding dimensionality D’ < D which doesn’t decrease embeddings correspondence quality (top-1 accuracy).
Neural MDS
We were studying the clusterization methods and show that there exist clusters in the embeddings data. Perform the visualization of these data clusters.
The first line contains the original images, the next two lines contain the modified ones. In this example, we change the components of the factorized vector and as a result we get photos of people with glasses.
By changing another component of the vector, the race of the people in the pictures changes.
By changing the component of the vector, we can also change the gender of a person
In the CLIP model, we can compare a text string with an image and determine how close they are to each other.
