Welcome to hybrid_learning’s documentation!
This is a pytorch implementation of parts of a hybrid learning lifecycle for computer vision convolutional neural networks (CNNs). It aims to verify that visual semantic concepts (defined by labeled examples) are (correctly) represented in the latent space of DNNs and correctly used. The core functionalities of the provided modules are:
Concept analysis: Finding and quality assessment of embeddings of concepts in a CNN latent space.
Model extension methods which allow to e.g. extend CNN outputs by concept predictions.
Custom dataset handles and some useful transformations for some standard concept datasets.
Logic framework: A framework to formulate, evaluate, and parse (fuzzy) logic rules, e.g., to check whether CNN outputs and concept predictions are plausible.
Experimentation utils for preparation, processing, and evaluation of standard experiments.