Out-of-distribution examples are images that are cearly irrelevant to the task at hand. Unfortunately, deep neural networks frequently assign random labels with high confidence to such examples. In this article, I want to discuss an adversarial way of computing high-confidence out-of-distribution examples, so-called distal adversarial examples, and how confidence-calibrated adversarial training handles them.
Properly evaluating defenses against adversarial examples has been difficult as adversarial attacks need to be adapted to each individual defense. This also holds for confidence-calibrated adversarial training, where robustness is obtained by rejecting adversarial examples based on their confidence. Thus, regular robustness metrics and attacks are not easily applicable. In this article, I want to discuss how to evaluate confidence-calibrated adversarial training in terms of metrics and attacks.
Recently, I had the opportunity to be a guest on Jay Shah’s podcast where he regularly talks to machine learning professionals from industry and academia. We had a great conversation about my PhD research and topics surrounding a successful career in machine learning — finding a good PhD program and research topic, preparing for interviews in industry, etc.
Taking adversarial training from this previous article as baseline, this article introduces a new, confidence-calibrated variant of adversarial training that addresses two significant flaws: First, trained with L∞ adversarial examples, adversarial training is not robust against L2 ones. Second, it incurs a significant increase in (clean) test error. Confidence-calibrated adversarial training addresses these problems by encouraging lower confidence on adversarial examples and subsequently rejecting them.
Knowing how to compute adversarial examples from this previous article, it would be ideal to train models for which such adversarial examples do not exist. This is the goal of developing adversarially robust training procedures. In this article, I want to describe a particularly popular approach called adversarial training. The idea is to train on adversarial examples computed during training on-the-fly. I will also discuss a PyTorch implementation that obtains 47.9% robust test error — 52.1% robust accuracy — on CIFAR10 using a WRN-28-10 architecture.
With our paper on conformal training, we showed how conformal prediction can be integrated into end-to-end training pipelines. There are so many interesting directions of how to improve and build upon conformal training. Unfortunately, I just do not have the bandwidth to pursue all of them. So, in this article, I want to share some research ideas so others can pick them up.
Adversarial examples, slightly perturbed images causing mis-classification, have received considerable attention over the last few years. While many different adversarial attacks have been proposed, projected gradient descent (PGD) and its variants is widely spread for reliable evaluation or adversarial training. In this article, I want to present my implementation of PGD to generate L∞, L2, L1 and L0 adversarial examples. Besides using several iterations and multiple attempts, the worst-case adversarial example across all iterations is returned and momentum as well as backtracking strengthen the attack.
Top-tier conferences in machine learning or computer vision generally require state-of-the-art results as baseline to assess novelty and significance of the paper. Unfortunately, getting state-of-the-art results on many benchmarks can be tricky and extremely time-consuming — even for rather simple benchmarks such as CIFAR-10. In this article, I want to share PyTorch code for obtaining 2.56% test error on CIFAR-10 using a Wide ResNet (WRN-28-10) and AutoAugment as well as Cutout for data augmentation.
PyTorch is a great tool to do deep learning research. However, when running large-scale experiments using various architectures, I always come across this one problem: How can I run the same experiments, evaluations or visualizations on models without knowing their architecture in advance? In this article, I want to present a simple approach allowing to load models without having to initialize the right architecture beforehand. The code of this article is available on GitHub.
Tensorboard is a great tool to monitor and debugg deep neural network training. Originally developed for TensorFlow, Tensorboard is now also supported by other libraries such as PyTorch. While the integration in PyTorch was shaky in the beginning, it got better and better with more recent releases. In this article, I want to discuss how to use Tensorboard for monitoring training with PyTorch. The article’s code is available on GitHub.