Optimize "Parametric UMAP" Tuned to MAX w/ AUTOENCODER

We code our Neural Networks (NN) for the Parametric UMAP encoder and decoder (like ConvNet for Vision MNIST data set) therefore we can optimize the performance of our parametric UMAP dimensional reduction algorithm. Plus activate the Autoencoder functionality for parametric UMAP.
The encoder is trained to minimize UMAP loss, and the decoder is trained to minimize reconstruction loss. To train the encoder jointly on both UMAP loss and reconstruction loss, pass "autoencoder_loss = True" into the ParametricUMAP.
We add specific loss function terms to balance local and global "nearness" of our graph and optimize for a minimal difference between our probability distributions, in the high-dim Input data space as well as in our embedding space, our low-dim target space.
original Jupyter NB:
github.com/lmcinnes/umap
github.com/lmcinnes/umap/tree...
All credits to:
arxiv.org/pdf/2009.12981.pdf
Tim Sainburg, Leland McInnes, Timothy Q Gentner
arxiv.org/pdf/1802.03426v1.pdf
Leland McInnes and John Healy
#datascience
#topologicalspace
#embedding
#umap
#pythonprogramming
#tensorflow2
#keras
@article{sainburg2021parametric,
title={Parametric UMAP Embeddings for Representation and Semisupervised Learning},
author={Sainburg, Tim and McInnes, Leland and Gentner, Timothy Q},
journal={Neural Computation},
volume={33},
number={11},
pages={2881--2907},
year={2021},
publisher={MIT Press One Rogers Street, Cambridge, MA 02142-1209, USA journals-info~…}
}

[@GentleAris]

Thanks for the informative video! Just a question, can we use 'reconstruction_loss' as a metric of how well our UMAP model has learned the data? What I want to say is, I suppose that an AE that has mapped our data well, has the ability to go from a low-dimensional space back to the high-dimensional space, by achieving the minimal loss. As a result, can we assume that the AE architecture that achieves the lowest 'reconstruction_losss' is the optimal one?