Overview In conformal training, a simple differentiable threshold conformal predictor is used during training in order to directly optimize losses defined on the predicted confidence sets. For example, let $E(x, k)$ be a conformity score, e.g., the logit or softmax probability for class $k$ of a classifier. Then, we can perform conformal prediction by defining the confidence sets as $C(x) := \{k: E(x, k) \geq \tau\}$ and calibrate the threshold $\tau$ on a calibration set $\{(x_i, y_i\}_{i = 1}^n$ using the $\alpha(1 + 1/n)$-quantile of the true scores $\{E(x_i,y_i)\}_{i = 1}^n$. This guarant ..read more

Introduction Since their discovery [], research on adversarial examples has exploded. Various different attacks have been proposed, but projected gradient descent (PGD) based attacks [][][] have become pretty standard in terms of robustness evaluation. Originally, the PGD attack was focussed on $L_\infty$-constrained adversarial examples. However, the concept can also be used to generate arbitrary $L_p$ adversarial examples such that improvements/changes of the PGD algorithm, in general, are widely applicable. In this article, I want to present my PyTorch implementation of projected gradient d ..read more

Introduction Computer vision benchmarks such as MNIST [] or CIFAR-10 [] are wide-spread within in research across various different task. However, independent of the task being tackled, it is nice — and often expected — to start with state-of-the-art performance as baseline. On MNIST, for example, this is rather easy to achieve and results are quite stable across neural network architectures and learning hyper-parameters. For CIFAR-10, in contrast, obtaining state-of-the-art performance can be time consuming and performance is more dependent on hyper-parameters. In this article, I want to shar ..read more

Introduction There is one aspect of deep learning frameworks such as PyTorch or TensorFlow that I find very cumbersome for researcher: Usually, only the model parameters are written to file, but not the architecture details. This means that, before loading a model from file, the architecture needs to be initialized. As result, I have to touch every training, evaluation or visualization script whenever adding a new architecture or merely changing hyper-parameters of an architecture. In this article, I want to present my approach to this problem. In particular, I want to be able to store and loa ..read more

Introduction For proper development and research of deep neural networks, tools such as TensorBoard for monitoring and debugging training are essential. Originally developed for TensorFlow, TensorBoard recently became a standalone package and support has been added to other deep learning frameworks such as PyTorch. As shown in Figure 1, Tensorboard has the advantage of plotting different data types without the need to implement the plotting itself. On the downside, the plots cannot easily be exported and integration with your favorit deep learning framework might be rather tedious and poorly d ..read more

Introduction Spell checking capabilities are getting better every year. While several companies offer good solutions for web browsers and good spell checking is built into most office applications, getting LaTeX projects/files spell-checked still seems problematic. While some of the web tools seem to work partly with OverLeaf, importing a LaTeX project into OverLeaf to manually spell-check all files is quite cumbersome. For me, GNU's aspell has long been to go-to tool to spell-check LaTeX files. For papers, this was usually a good solution as few files were involved and spell-checking only hap ..read more

Introduction Preparing a paper for ArXiv usually involves the same steps. Authors generally want to remove comments and unused files — for example, images for figures/plots. Also, ArXiv usually expects a "flat" submission, that is, ArXiv does not work well with sub-directories in the uploaded submission. Over the years, I came up with a set of Python scripts to remove unused files and flatten the remaining files. In this article, I want to share these scripts. The code including basic documentation can be found on GitHub: Code on GitHub Requirements are minimal and should not require any Pytho ..read more

Introduction Most PhD programs share the same goal of performing original academic research. In this sense, a PhD can be seen as a qualification to perform independent research and is usually required to work as a professor at a university or a researcher in general. In many disciplines, a PhD involves publications at peer-reviewed journals and/or conferences. This is generally the publicly visible part of a PhD: Publications are not only easily accessible online but often also involve talks or posters at the corresponding conferences. Despite this common goal of conducting original research ..read more

Each year, thousands of students decide to pursue a PhD. Especially in artifical intelligence, PhD programs have been incredibly popular in recent years. However, PhD programs can vary significantly across countries, institutions and advisors. Personally, it took me roughly 2-3 years to actually realize that a PhD is the right path for me. These years were filled with a master degree, several industry internships and varying experiences in academia. Thus, when it came to choosing where and with whom to do my PhD, I had the advantage to know how PhD programs in the United States look, how an in ..read more

Thesis Figure 1: Problems tackled in my PhD thesis, ranging from 3D reconstruction in the context of autonomous driving to adversarial robustness, bit error robustness and uncertainty estimation. The thesis can be downloaded from Saarland University's library — the Saarländische Universitäts- und Landesbibliothek (SULB). Find the abstract, table of contents and included publications below. The thesis is roughly 17MB of size: PhD Thesis PDF @phdthesis{Stutz2022, title={Understanding and improving robustness and uncertainty estimation in deep learning}, author={David Stutz}, scho ..read more