Posts by Tags

1-Lipschitz

How Many Power Iteration Steps do 1-Lipschitz Networks Need?

12 minute read

Published:

Power iteration is a widely used algorithm for constructing 1-Lipschitz layers. In this blog post we explore how many steps are required to obtain reliable guarantees. The takeaway? It’s far more than most papers assume. Power Iteration Convergence

Hard Mining for Robust Classification

6 minute read

Published:

In this blog post we want to explore whether training mostly on the hardest examples allows us to fit robust networks on CIFAR-10 quicker.

Adversarial Robustness

Are LMMs Robust to Small Image Perturbations?

6 minute read

Published:

In this blog post we investigate whether Large Multimodal Models (LMMs) are robust to small perturbations of input images. We find that this is not the case, we manage to fool Phi-4-multimodal-instruct by small changes to its inputs. LMM Robustness

Certified Robust Classification

Hard Mining for Robust Classification

6 minute read

Published:

In this blog post we want to explore whether training mostly on the hardest examples allows us to fit robust networks on CIFAR-10 quicker.

Clean Code

Elegant Configurations in Python

5 minute read

Published:

In my python code I always struggled to define a configuration in a way that is easily readable, easy to understand and allows auto-complete and changes from the command line. This blog post is an attempt to do exactly that.

Configuration

Elegant Configurations in Python

5 minute read

Published:

In my python code I always struggled to define a configuration in a way that is easily readable, easy to understand and allows auto-complete and changes from the command line. This blog post is an attempt to do exactly that.

Hard Mining

Hard Mining for Robust Classification

6 minute read

Published:

In this blog post we want to explore whether training mostly on the hardest examples allows us to fit robust networks on CIFAR-10 quicker.

Hyperparameter

Elegant Configurations in Python

5 minute read

Published:

In my python code I always struggled to define a configuration in a way that is easily readable, easy to understand and allows auto-complete and changes from the command line. This blog post is an attempt to do exactly that.

LLM

Are LMMs Robust to Small Image Perturbations?

6 minute read

Published:

In this blog post we investigate whether Large Multimodal Models (LMMs) are robust to small perturbations of input images. We find that this is not the case, we manage to fool Phi-4-multimodal-instruct by small changes to its inputs. LMM Robustness

Phi 4

Are LMMs Robust to Small Image Perturbations?

6 minute read

Published:

In this blog post we investigate whether Large Multimodal Models (LMMs) are robust to small perturbations of input images. We find that this is not the case, we manage to fool Phi-4-multimodal-instruct by small changes to its inputs. LMM Robustness

Power Iteration

How Many Power Iteration Steps do 1-Lipschitz Networks Need?

12 minute read

Published:

Power iteration is a widely used algorithm for constructing 1-Lipschitz layers. In this blog post we explore how many steps are required to obtain reliable guarantees. The takeaway? It’s far more than most papers assume. Power Iteration Convergence

Priority Queue

Hard Mining for Robust Classification

6 minute read

Published:

In this blog post we want to explore whether training mostly on the hardest examples allows us to fit robust networks on CIFAR-10 quicker.