Publications

NeurIPS 2023 (Spotlight)

Scale Alone Does not Improve Mechanistic Interpretability in Vision Models

Zimmermann*, R. S., Klein*, T. and Brendel, W.

We compare the mechanistic interpretability of vision models differing with respect to scale, architecture, training paradigm and dataset size and find that none of these design choices have any significant effect on the interpretability of individual units. We release a dataset of unit-wise interpretability scores that enables research on automated alignment.

Access the paper

Preprint

Don’t trust your eyes: on the (un)reliability of feature visualizations

Geirhos*, R., Zimmermann*, R. S., Bilodeau*, B., Brendel, W., Kim, B

Feature visualizations seek to explain how neural networks process natural images, but as we show both experimentally and analytically, they can be unreliable (for instance, they can be manipulated to show arbitrary patterns).

Access the paper

Preprint

Sensitivity of Slot-Based Object-Centric Models to their Number of Slots

Zimmermann, R. S., van Steenkiste, S., Sajjadi, M. S. M., Kipf, T., Greff, K.

We investigate the sensitivity of slot-based object-centric models to their number of slots, identify failures and explore mitigation strategies.

Access the paper

ICML 2023 (Oral)

Provably Learning Object-Centric Representations

Brady*, J., Zimmermann*, R. S., Sharma*, Y., Schölkopf, B., von Kügelken, J. and Brendel, W.,

We analyze when object-centric representations can be learned without supervision and introduces two assumptions, compositionality and irreducibility, to prove that ground-truth object representations can be identified.

Access the paper

NeurIPS 2022

Increasing Confidence in Adversarial Robustness Evaluations

Zimmermann, R. S., Brendel, W., Tramer, F., and Carlini, N.

We propose a test that enables researchers to find flawed adversarial robustness evaluations. Passing our test produces compelling evidence that the attacks used have sufficient power to evaluate the model’s robustness.

Access the paper

NeurIPS 2021 (Spotlight)

How Well do Feature Visualizations Support Causal Understanding of CNN Activations?

Zimmermann*, R. S., Borowski*, J., Geirhos, R. , Bethge, M., Wallis, T. S. A. and Brendel, W.

Using psychophysical experiments, we show that state-of-the-art synthetic feature visualizations do not support causal understanding much better than no visualizations, and only similarly well as other visualizations like natural dataset samples.

Access the paper

workshop

Score-Based Generative Classifiers

Zimmermann, R. S., Schott, L., Song, Y. Dunn, B. A. and Klindt, D. A.

We evaluate score-based generative models as classifiers on CIFAR-10 and find that they yield good accuracy and likelihoods but no adversarial robustness.

Access the workshop paper

article

Reconstructing complex cardiac excitation waves from incomplete data using echo state networks and convolutional autoencoders

Herzog, S., Zimmermann, R. S., Abele, J., Luther, S. and Parlitz, U.

We use two approaches from machine learning, echo state networks and convolutional autoencoders, to solve two relevant data modelling tasks in cardiac dynamics to bridge the gap between measurable and not measurable quantities

Access the paper

ICLR 2021

Exemplary Natural Images Explain CNN Activations Better than State-of-the-Art Feature Visualization

Borowski*, J., Zimmermann*, R. S., Schepers, J., Geirhos, R. , Wallis, T. S. A., Bethge, M. and Brendel, W.

Using human psychophysical experiments, we show that natural images can be significantly more informative for interpreting neural network activations than synthetic feature visualizations.

Access the paper

ICML 2021 (Spotlight)

Contrastive Learning Inverts the Data Generating Process

Zimmermann*, R. S., Sharma*, Y., Schneider*, S., Bethge, M. and Brendel, W.,

We rigorously show that models trained on a common contrastive loss can implicitly invert the underlying generative model of the observed data up to affine transformations.

Access the paper

workshop paper

Contrastive Learning can Identify the Underlying Generative Factors of the Data

Zimmermann*, R. S., Schneider*, S. Sharma*, Y., Bethge, M. and Brendel, W.,

We rigorously show that models trained on a common contrastive loss can implicitly invert the underlying generative model of the observed data up to affine transformations.

Access the workshop paper

workshop paper

Natural images are more informative for interpreting CNN activations than synthetic feature visualizations

Borowski*, J., Zimmermann*, R. S., Schepers, J., Geirhos, R. , Wallis, T. S. A., Bethge, M. and Brendel, W.

Using human psychophysical experiments, we show that natural images can be significantly more informative for interpreting neural network activations than synthetic feature visualizations.

Access the workshop paper

paper

Foolbox Native: Fast adversarial attacks to benchmark the robustness of machine learning models in PyTorch, TensorFlow, and JAX

Jonas, R., Zimmermann, R. S., Bethge, M. and Brendel, W.,

Foolbox is a popular Python library to benchmark the robustness of machine learning models against these adversarial perturbations

Access the workshop paper

ECCV 2020 (Oral)

A simple way to make neural networks robust against diverse image corruptions

Rusak*, E. and Schott*, L. and Zimmermann*, R. S. and Bitterwolf, J. and Bringmann, O. and Bethge, M. and Brendel, W.

Simple but properly tuned training with additive noise generalizes surprisingly well to unseen corruptions.

Access the paper

workshop paper

Increasing the robustness of DNNs against image corruptions by playing the Game of Noise

Rusak*, E. and Schott*, L. and Zimmermann*, R. S. and Bitterwolf, J. and Bringmann, O. and Bethge, M. and Brendel, W.

Simple but properly tuned training with additive noise generalizes surprisingly well to unseen corruptions.

Access the workshop paper

thesis

Robust Perception in Humans and Machines

Zimmermann, R. S.

Analyzing the robustness of computer vision and comparing it to human performance by developing novel evaluation methods

Access the thesis

comment

Comment on “Adv-BNN: Improved Adversarial Defense through Robust Bayesian Neural Network”

Zimmermann, R. S.

Breaks the the proposed adversarial defense method Adv-BNN

Access the comment

poster

Predicting and Observing Chaotic Dynamics in Excitable Media Using Machine Learning

Parlitz, U. and Zimmermann, R. S. , Herzog, S. and Isenseee, J. and Datseris, G.

Analysis of chaotic dynamics using sequential and recurrent neural networks and classical Machine Learning methods.

article

Observing spatio-temporal dynamics of excitable media using reservoir computing

Zimmermann, R. S. and Parlitz, U.

Examination of recurrent neural networks (Echo State Networks) for the spatio-temporal cross prediction of chaotic systems.

Access the paper

preprint

Faster Training of Mask R-CNN by Focusing on Instance Boundaries

Zimmermann*, R. S. and Siems*, J. N.

Improving the training of Mask R-CNN for instance segmentation by introducing an intuitive auxiliary loss.

Access the preprint

thesis

Prediction spatio-temporal dynamics using reservoir computing

Zimmermann, R. S.

Examination of recurrent neural networks (Echo State Networks) for the multiple prediction tasks in chaotic systems.

Access the thesis