AI Projects & Research
My hobbyist research projects, art projects, and thesis
MS Thesis: Visualizing and comparing standardized Neural Network Encodings using Manifold Learning
The thesis describes a method to visualize the representation of a system that a neural network has learned in a standardized and comparable way using manifold learning in the form of Diffusion Maps with the Mahalanobis metric and Autoencoders.
Developed a novel method using NCAs to transform images iteratively for style transfer and text-guided image transformations using a CLIP-based loss function. The method application yields good results and is applicable in real-time applications and due to the nature of NCAs is robust to overfitting.
A transformer embedding-based search engine for around 100.000 forums on ”Reddit”. This service was made to address the lack of a good recommendation service offered by Reddit itself.
CLIP-based rating of aesthetics in Image Regions
Using @Rivershavewings regression of aesthetics rating in CLIP embedding space a method was constructed to perform a patch-wise rating of an image’s aesthetics. This results in a map of aesthetics as judged by the Model, which was trained on human ratings, showing the regions of high and low aesthetics. This method can help illustrators and artists to evaluate their work or create masks for inpainting using diffusion Models automatically.
