Taking the TensorBoard Embedding Projector to the Next Level

The TensorBoard embedding projector is a very powerful tool in data analysis, specifically for interpreting and visualizing low-dimensional embeddings. In order to do so, first, it applies a dimensionality reduction algorithm to the input embeddings, between UMAP, T-SNE, PCA, or a custom one, to reduce their dimension to three and be able to render them in a three-dimensional space. Once the map is generated, this tool can be used, for example, to search for specific keywords associated with the embeddings or highlight similar points in space. Ultimately, its goal is to provide a way to better interpret the embeddings that our machine learning model is generating, to check if the similar ones according to our definition are plotted nearby in the 3D space. When the embeddings we want to display originate from words, plotting the points with the tag it refers is more than enough, which is exactly the use case presented in the TensorBoard documentation.