Deep Learning Techniques

We take a multi-disciplinary perspective to develop inductive biases to increase generalisation capabilities of deep neural networks. Particularly, we invent methods which induce and exploit structure of the data, the model, and its training.

Currently, our main strategies focus on exploiting invariances and finding interesting projects for a given task (Deep Projection Pursuit).

Deep Projection Pursuit

Deep Projection Pursuit deals with finding interesting projections to understand data or solve a given task. For that, we consider models with three steps:

  1. Input: Prepare inputs
  2. Projections: Determine interesting non-linear projections
  3. Shape Functions: Solve a machine learning task given the parameters of the projection or the projected inputs
  4. Aggregation: Summarize the shape function values for the task

This general framework allows developing methods for different steps separately which lead to the development of different approaches for projections such as quantization approaches or VAE-based approaches. Moreover, we developed full implementation of the framework called INAM which allows performing interpretable image classification.

Publications

  • J. Hüls, J.Perschewski, S.Stober. INAM: Image-Scale Neural Additive Models.European Symposium on Artificial Neural Networks, Computational Intelligence and Machine Learning, 2025.[Poster]
  • J. Perschewski, J. Schmidt, S. Stober. Pursuing the Perfect Projection: A Projection Pursuit Framework for Deep Learning. International Workshop on Self-Organizing Maps and Learning Vector Quantization, Clustering and Data Visualization (WSOM+), 2024. [Link]
  • J.Perschewski, S.Stober.T-DVAE: A Transformer-Based Dynamical Variational Autoencoder for Speech. International Conference on Artificial Neual Networks, 2024.  [Link]
  • J.Perschewski, S.Stober.Neural-Gas VAE. International Conference on Artificial Neual Networks, 2022 [Link]

 

Last Modification: 13.05.2025 -
Contact Person: Webmaster