direkt zum Inhalt springen

direkt zum Hauptnavigationsmenü

Sie sind hier

TU Berlin

Inhalt des Dokuments

Grégoire Montavon

Lupe

Biography

Grégoire Montavon is a senior researcher in the Machine Learning Group at the Technische Universität Berlin, and in the Berlin Institute for the Foundations of Learning and Data (BIFOLD). He received a Masters degree in Communication Systems from École Polytechnique Fédérale de Lausanne in 2009, and a Ph.D. degree in Machine Learning from the Technische Universität Berlin in 2013. His research interests include explainable AI, deep neural networks, and unsupervised learning.

 

 

Research on Explainable AI

Lupe

The goal of this project is to develop methods to explain the decisions (e.g. classifications) of machine learning models in terms of input variables (i.e. which input variables are relevant for the model's decision).

  • Explanations in deep neural networks is particularly difficult due to the high level of nonlinearity of the decision function. We have proposed a procedure called Layer-Wise Relevance Propagation (LRP) that robustly explains deep neural networks at low computational cost.
  • We have proposed the Deep Taylor Decomposition framework, which connects mathematically the LRP procedure to Taylor expansions. Specifically, the LRP algorithm can be interpreted as a collection of Taylor expansions performed locally in the network, and this connection allows us to systematically bring LRP to new architectures with different layer types.
  • In order to apply explanation techniques beyond neural networks, we have proposed a Neuralization-Propagation approach which consists of (1) rewriting non neural network models such as kernel-based anomaly detection or k-means clustering as strictly equivalent neural networks, and (2) use the neural network representation and LRP to produce the explanation. Unlike surrogate modeling approaches, our neuralization-propagation approach does not require any retraining.
  • Because a simple attribution of the prediction on the input features may not be sufficient for certain tasks (e.g. predicting similarity or predicting graphs), we have developed extensions for these cases: BiLRP and GNN-LRP, which produce explanations in terms of pairs of input features or in terms of walks into the graph.
  • The methods we have developed can serve to verify that trained neural networks predict as expected (i.e. are not subject to a Clever Hans effect), and we are also starting to apply our method to address challenges in the area of digital humanities and quantum chemistry.

Demos, tutorials, software, and list of publications are available at www.heatmapping.org

Teaching

Winter semester 2020/2021:

Summer semester 2020:

Publications

Preprints

 Edited books

 Book chapters

Journal Publications

Conference Publications

Software Publications

Workshop Publications

Theses

Unpublished

Zusatzinformationen / Extras

Quick Access:

Schnellnavigation zur Seite über Nummerneingabe

Auxiliary Functions