Hack and explore the latent space
Hack and explore the latent space
Not a complete challenge anymore, but under the umbrella of Hack the planets or Hack the stellar populations
Category: Unsupervised and self-supervised learning, visual analytics, clustering
Your challenge is to show nontrivial groupings of images in a (potentially multispectral) visual dataset of planetary surface images or astrophysical images by learning features in an unsupervised way. You can use a Python-based notebook we will provide to extract information from your dataset, or you could implement your own code/method. Mandatory: use the Latent Space Explorer tool for sake of understanding the task at hand and comparing results.
Sub-Tasks
Choose an image dataset: it could be formed by standard RGB images or multi-spectral images. (hint: use a multi-labeled dataset)
Train a model that extracts features (latent vectors), example:
Convolutional AutoEncoder (provided)
SimClr (provided)
One of your choice/implementation (not provided)
Load results of your experiments into NEANIAS NextCloud
Explore the representation space using Latent Space Explorer service
Reduce the dimensionality of the representation by using pca/tsne/umap (or implement on your PC an algorithm of your choice and visualize results)
Perform clustering algorithm on the dataset using one of the pre-provided algorithms or implement on your PC an algorithm of your choice
If present, visualize the ground-truth labels
Provide some metrics (confusion matrix, if ground truth is available) or qualitative observations about the results
Background
AIxIA Discussion Paper – upcoming
Google TensorFlow Blog: Introducing TensorFlow Similarity
Indicative Data Sources
Usual suspects, increasing complexity and difficulty in evaluating quantitatively achieved results:
CelebA dataset (subsets of it)
Multispectral Image Recognition (subsets of it)