Márk Jelasity

We are concerned with the problem of identifying samples with noisy labels in a given dataset. Using the predictions of well-generalizing models to flag incorrectly predicted labels as noisy is a known method, but it is generally not considered competitive. At the same time, it has been observed recently that gradient descent fits clean samples first, and the noisy samples are memorized later. Inspired by related theoretical results, we revisit the idea of using the predictions of well-generalizing model checkpoints during training to classify samples as noisy or clean. We offer two key improvements that allow this strikingly simple approach to outperform some well-known methods. Our first method, ES-Prediction, classifies a sample as noisy if the model checkpoint predicts its label incorrectly. Our second method, ES-BGSS, is based on the observation that the sum of the per-sample gradients of the loss with respect to the feature layer bias parameters tends to be negative for clean samples and positive for noisy ones. Both methods are applied to an ensemble of model checkpoints collected around the early stopping point of a standard training run, which reduces the variance of the predictions and improves detection accuracy. Both methods operate in-sample, meaning they evaluate each checkpoint on its own training set and directly exploit the clean priority learning phenomenon. Our approaches only passively observe a normal training run and collect checkpoints. No extra input samples are added, and no pretrained models are used. Through empirical evaluations, we demonstrate that our methods are competitive with other approaches from related work for both detecting noisy samples and for noise-filtered training.

Generative sequence models are typically trained on sample sequences from natural or formal languages. It is a crucial question whether -- or to what extent -- sample-based training is able to capture the true structure of these languages, often referred to as the ``world model''. Theoretical results indicate that we can hope for soundness at best, that is, generating valid sequences, but not necessarily all of them. However, it is still important to have practical tools that are able to verify whether a given sequence model is sound. In this study, we focus on chess, as it is a domain that provides enough complexity while having a simple rule-based world model. We propose adversarial sequence generation for verifying the soundness of the sequence model. Our adversaries generate valid sequences so as to force the sequence model to generate an invalid next move prediction. Apart from the falsification of soundness, this method is also suitable for a more fine-grained analysis of the failure modes and the effects of different choices during training. To demonstrate this, we propose a number of methods for adversarial sequence generation and evaluate the approach on a large set of chess models. We train models on random as well as high-quality chess games, using several training recipes. We find that none of the models are sound, but some training techniques and dataset choices are able to improve soundness remarkably. We also investigate the potential application of board state probes in both our training and attack methods. Our findings indicate that the extracted board states have no causal role in next token prediction in most of the models.

Detecting semantic backdoors in classification models--where some classes can be activated by certain natural, but out-of-distribution inputs--is an important problem that has received relatively little attention. Semantic backdoors are significantly harder to detect than backdoors that are based on trigger patterns due to the lack of such clearly identifiable patterns. We tackle this problem under the assumption that the clean training dataset and the training recipe of the model are both known. These assumptions are motivated by a consumer protection scenario, in which the responsible authority performs mystery shopping to test a machine learning service provider. In this scenario, the authority uses the provider's resources and tools to train a model on a given dataset and tests whether the provider included a backdoor. In our proposed approach, the authority creates a reference model pool by training a small number of clean and poisoned models using trusted infrastructure, and calibrates a model distance threshold to identify clean models. We propose and experimentally analyze a number of approaches to compute model distances and we also test a scenario where the provider performs an adaptive attack to avoid detection. The most reliable method is based on requesting adversarial training from the provider. The model distance is best measured using a set of input samples generated by inverting the models in such a way as to maximize the distance from clean samples. With these settings, our method can often completely separate clean and poisoned models, and it proves to be superior to state-of-the-art backdoor detectors as well.

Measuring the similarity of the internal representations of deep neural networks is an important and challenging problem. Model stitching has been proposed as a possible approach, where two half-networks are connected by mapping the output of the first half-network to the input of the second one. The representations are considered functionally similar if the resulting stitched network achieves good task-specific performance. The mapping is normally created by training an affine stitching layer on the task at hand while freezing the two half-networks, a method called task loss matching. Here, we argue that task loss matching may be very misleading as a similarity index. For example, it can indicate very high similarity between very distant layers, whose representations are known to have different functional properties. Moreover, it can indicate very distant layers to be more similar than architecturally corresponding layers. Even more surprisingly, when comparing layers within the same network, task loss matching often indicates that some layers are more similar to a layer than itself. We argue that the main reason behind these problems is that task loss matching tends to create out-of-distribution representations to improve task-specific performance. We demonstrate that direct matching (when the mapping minimizes the distance between the stitched representations) does not suffer from these problems. We compare task loss matching, direct matching, and well-known similarity indices such as CCA and CKA. We conclude that direct matching strikes a good balance between the structural and functional requirements for a good similarity index.

Measuring the similarity of the internal representations of deep neural networks is an important and challenging problem. Model stitching has been proposed as a possible approach, where two half-networks are connected by mapping the output of the first half-network to the input of the second one. The representations are considered functionally similar if the resulting stitched network achieves good task-specific performance. The mapping is normally created by training an affine stitching layer on the task at hand while freezing the two half-networks, a method called task loss matching. Here, we argue that task loss matching may be very misleading as a similarity index. For example, it can indicate very high similarity between very distant layers, whose representations are known to have different functional properties. Moreover, it can indicate very distant layers to be more similar than architecturally corresponding layers. Even more surprisingly, when comparing layers within the same network, task loss matching often indicates that some layers are more similar to a layer than itself. We argue that the main reason behind these problems is that task loss matching tends to create out-of-distribution representations to improve task-specific performance. We demonstrate that direct matching (when the mapping minimizes the distance between the stitched representations) does not suffer from these problems. We compare task loss matching, direct matching, and well-known similarity indices such as CCA and CKA. We conclude that direct matching strikes a good balance between the structural and functional requirements for a good similarity index.

Machine learning models are vulnerable to tiny adversarial input perturbations optimized to cause a very large output error. To measure this vulnerability, we need reliable methods that can find such adversarial perturbations. For image classification models, evaluation methodologies have emerged that have stood the test of time. However, we argue that in the area of semantic segmentation, a good approximation of the sensitivity to adversarial perturbations requires significantly more effort than what is currently considered satisfactory. To support this claim, we re-evaluate a number of well-known robust segmentation models in an extensive empirical study. We propose new attacks and combine them with the strongest attacks available in the literature. We also analyze the sensitivity of the models in fine detail. The results indicate that most of the state-of-the-art models have a dramatically larger sensitivity to adversarial perturbations than previously reported. We also demonstrate a size-bias: small objects are often more easily attacked, even if the large objects are robust, a phenomenon not revealed by current evaluation metrics. Our results also demonstrate that a diverse set of strong attacks is necessary, because different models are often vulnerable to different attacks.

Achieving robustness against adversarial input perturbation is an important and intriguing problem in machine learning. In the area of semantic image segmentation, a number of adversarial training approaches have been proposed as a defense against adversarial perturbation, but the methodology of evaluating the robustness of the models is still lacking, compared to image classification. Here, we demonstrate that, just like in image classification, it is important to evaluate the models over several different and hard attacks. We propose a set of gradient based iterative attacks and show that it is essential to perform a large number of iterations. We include attacks against the internal representations of the models as well. We apply two types of attacks: maximizing the error with a bounded perturbation, and minimizing the perturbation for a given level of error. Using this set of attacks, we show for the first time that a number of models in previous work that are claimed to be robust are in fact not robust at all. We then evaluate simple adversarial training algorithms that produce reasonably robust models even under our set of strong attacks. Our results indicate that a key design decision to achieve any robustness is to use only adversarial examples during training. However, this introduces a trade-off between robustness and accuracy.