Path attribution methods are a gradient-based way of explaining deep models. These methods require choosing a hyperparameter known as the baseline input. What does this hyperparameter mean, and how important is it? In this article, we investigate these questions using image classification networks as a case study. We discuss several different ways to choose a […]

While deep neural networks have overwhelmingly established state-of-the-art results in many artificial intelligence problems, they can still be difficult to develop and debug. Recent research on deep learning understanding has focused on feature visualization , theoretical guarantees , model interpretability , and generalization . In this work, we analyze deep neural networks from a complementary

Introduction In the last few years, reinforcement learning (RL) has made remarkable progress, including beating world-champion Go players, controlling robotic hands, and even painting pictures. One of the key sub-problems of RL is value estimation – learning the long-term consequences of being in a state. This can be tricky because future returns are generally noisy,

Section 3.2 of Ilyas et al. (2019) shows that training a model on only adversarial errors leads to non-trivial generalization on the original test set. We show that these experiments are a specific case of learning from errors. We start with a counterintuitive result — we take a completely mislabeled training set (without modifying the inputs) and

We demonstrate that there exist adversarial examples which are just “bugs”: aberrations in the classifier that are not intrinsic properties of the data distribution. In particular, we give a new method for constructing adversarial examples which: Do not transfer between models, and Do not leak “non-robust features” which allow for learning, in the sense of

A figure in Ilyas, et. al. that struck me as particularly interesting was the following graph showing a correlation between adversarial transferability between architectures and their tendency to learn similar non-robust features. Adversarial transferability vs test accuracy of different architectures trained on ResNet-50′s non-robust features. One way to interpret this graph is that it shows

A Discussion of ‘Adversarial Examples Are Not Bugs, They Are Features’: Two Examples of Useful, Non-Robust Features Ilyas et al. define a feature as a function fff that takes xxx from the data distribution (x,y)∼D(x,y) \sim \mathcal{D}(x,y)∼D into a real number, restricted to have mean zero and unit variance. A feature is said to be

Ilyas et al. report a surprising result: a model trained on adversarial examples is effective on clean data. They suggest this transfer is driven by adverserial examples containing geuinely useful non-robust cues. But an alternate mechanism for the transfer could be a kind of “robust feature leakage” where the model picks up on faint robust

The hypothesis in Ilyas et. al. is a special case of a more general principle that is well accepted in the distributional robustness literature — models lack robustness to distribution shift because they latch onto superficial correlations in the data. Naturally, the same principle also explains adversarial examples because they arise from a worst-case analysis of distribution

We want to thank all the commenters for the discussion and for spending time designing experiments analyzing, replicating, and expanding upon our results. These comments helped us further refine our understanding of adversarial examples (e.g., by visualizing useful non-robust features or illustrating how robust models are successful at downstream tasks), but also highlighted aspects of