Interests
- Human vs Machine Comparisons
- Generalization
- Robustness
- Identifiability
- Internal World Model Building
- Sample Efficiency
- Vision-Language-Action Models
Education
Ph.D. in Computational Neuroscience, 2014
École Normale Supérieure
Diploma (M.Sc. equivalent) in Physics, 2010
University of Regensburg
Latest
- LLMs on the Line: Data Determines Loss-to-Loss Scaling Laws
- Pretraining Frequency Predicts Compositional Generalization of CLIP on Real-World Tasks
- Interaction Asymmetry: A General Principle for Learning Composable Abstractions
- Cross-Entropy Is All You Need To Invert the Data Generating Process
- In search of forgotten domain generalization
- Error consistency between humans and machines as a function of presentation duration
- Rule Extrapolation in Language Models: A Study of Compositional Generalization on OOD Prompts
- InfoNCE: Identifying the Gap Between Theory and Practice
- Identifiable Exchangeable Mechanisms for Causal Structure and Representation Learning
- Translational symmetry in convolutions with localized kernels causes an implicit bias toward high frequency adversarial examples
- Position: Understanding LLMs requires more than statistical generalization
- Understanding LLMs Requires More Than Statistical Generalization
- An interventional perspective on identifiability in gaussian lti systems with independent component analysis
- Effective pruning of web-scale datasets based on complexity of concept clusters
- Does CLIP's Generalization Performance Mainly Stem from High Train-Test Similarity?
- Provable Compositional Generalization for Object-Centric Learning
- Feature Visualizations do not sufficiently explain hidden units of Artificial Neural Networks
- Scale Alone Does not Improve Mechanistic Interpretability in Vision Models
- Compositional Generalization from First Principles
- Don't trust your eyes: on the (un) reliability of feature visualizations
- Provably Learning Object-Centric Representations
- Robust deep learning object recognition models rely on low frequency information in natural images
- Iterative weakly supervised learning for novel class object detection
- Jacobian-based Causal Discovery with Nonlinear ICA
- Learning From Brains How to Regularize Machines (Supplementary Material)
- The Independent Compositional Subspace Hypothesis for the Structure of CLIP's Last Layer
- The bittersweet lesson: data-rich models narrow the behavioural gap to human vision
- If your data distribution shifts, use self-learning
- Increasing Confidence in Adversarial Robustness Evaluations
- Embrace the Gap: VAEs Perform Independent Mechanism Analysis
- Content suppresses style: dimensionality collapse in contrastive learning
- Imagenet-d: A new challenging robustness dataset inspired by domain adaptation
- 2.1 Decision-Based Adversarial Attacks
- Visual representation learning does not generalize strongly within the same domain
- How Well do Feature Visualizations Support Causal Understanding of CNN Activations?
- Partial success in closing the gap between human and machine vision
- Self-Supervised Learning with Data Augmentations Provably Isolates Content from Style
- Adapting imagenet-scale models to complex distribution shifts with self-learning
- Five points to check when comparing visual perception in humans and machines
- Fast Minimum-norm Adversarial Attacks through Adaptive Norm Constraints
- Contrastive Learning Inverts the Data Generating Process
- Exemplary Natural Images Explain CNN Activations Better than Feature Visualizations
- Unintended cue learning: Lessons for deep learning from experimental psychology
- On the surprising similarities between supervised and self-supervised models
- Foolbox native: Fast adversarial attacks to benchmark the robustness of machine learning models in pytorch, tensorflow, and jax
- EagerPy: Writing code that works natively with PyTorch, TensorFlow, JAX, and NumPy
- Towards Nonlinear Disentanglement in Natural Data with Temporal Sparse Coding
- Improving robustness against common corruptions by covariate shift adaptation
- Local convolutions cause an implicit bias towards high frequency adversarial examples
- Unmasking the inductive biases of unsupervised object representations for video sequences
- Local Convolutions Cause an Implicit Bias towards High Frequency Adversarial Examples
- Shortcut Learning in Deep Neural Networks
- Learning to represent signals spike by spike
- On adaptive attacks to adversarial example defenses
- Increasing the robustness of DNNs against image corruptions by playing the Game of Noise
- A simple way to make neural networks robust against diverse image corruptions
- Adversarial vision challenge
- Generalized Invariant Risk Minimization: relating adaptation and invariant representation learning
- Inducing a human-like shape bias leads to emergent human-level distortion robustness in CNNs
- Benchmarking robustness in object detection: Autonomous driving when winter is coming
- Reproducing Decision-Making With Constrained Networks to Understand Deep Neural Networks
- Approximating cnns with bag-of-local-features models works surprisingly well on imagenet
- On evaluating adversarial robustness
- Accurate, reliable and fast robustness evaluation
- Learning from brains how to regularize machines
- ImageNet-trained CNNs are biased towards texture; increasing shape bias improves accuracy and robustness
- Comparing the ability of humans and DNNs to recognise closed contours in cluttered images
- Biomechanical texture coding in rat whiskers
- One-shot texture segmentation
- Towards the first adversarially robust neural network model on MNIST
- Trace your sources in large-scale data: one ring to find them all
- Decision-based adversarial attacks: Reliable attacks against black-box machine learning models
- Foolbox v0. 8.0: A python toolbox to benchmark the robustness of machine learning models
- Comment on" Biologically inspired protection of deep networks from adversarial attacks"
- What does it take to generate natural textures?
- Texture synthesis using shallow convolutional networks with random filters
- Demixed principal component analysis of neural population data
- On the demixing and self-organized formation of neural population responses
- Unsupervised learning of an efficient short-term memory network
- Demixed principal component analysis
- Covariant boost and structure functions of baryons in Gross-Neveu models
- Instanton constituents and fermionic zero modes in twisted CPn models