Adversarial training (AT) is one of the most effective strategies for promoting model robustness. However, recent benchmarks show that most of the proposed improvements on AT are less effective than simply early stopping the training procedure. This counter-intuitive fact motivates us to investigate the implementation details of tens of AT methods. Surprisingly, we find that the basic training settings (e.g., weight decay, learning rate schedule, etc.) used in these methods are highly inconsistent, which could largely affect the model performance as shown in our experiments. For example, a slightly different value of weight decay can reduce the model robust accuracy by more than 7%, which is probable to override the potential promotion induced by the proposed methods. In this work, we provide comprehensive evaluations on the effects of basic training tricks and hyperparameter settings for adversarially trained models. We provide a reasonable baseline setting and re-implement previous defenses to achieve new state-of-the-art results.
Recent successes of Deep Neural Networks (DNNs) in a variety of research tasks, however, heavily rely on the large amounts of labeled samples. This may require considerable annotation cost in real-world applications. Fortunately, active learning is a promising methodology to train high-performing model with minimal annotation cost. In the deep learning context, the critical question of active learning is how to precisely identify the informativeness of samples for DNN. In this paper, inspired by piece-wise linear interpretability in DNN, we firstly introduce the linear separable regions of samples to the problem of active learning, and propose a novel Deep Active learning approach by Model Interpretability (DAMI). To keep the maximal representativeness of the entire unlabeled data, DAMI tries to select and label samples on different linear separable regions introduced by the piece-wise linear interpretability in DNN. We focus on two scenarios: 1) Multi-Layer Perception (MLP) for modeling tabular data; 2) language models for modeling textual data. On tabular data, we use the local piece-wise interpretation in DNN as the representation of each sample, and directly run K-Center clustering to select and label the central sample in each cluster. On textual data, we propose a novel aggregator to find the most informative word in each sentence, and use its local piece-wise interpretation as the representation of the sentence. To be noted, this whole process of DAMI does not require any hyper-parameters to tune manually. To verify the effectiveness of our approach, extensive experiments have been conducted on both tabular datasets and textual datasets. The experimental results demonstrate that DAMI constantly outperforms several state-of-the-art compared methods.
Convolutional neural networks (CNNs) have been successfully used in a range of tasks. However, CNNs are often viewed as "black-box" and lack of interpretability. One main reason is due to the filter-class entanglement -- an intricate many-to-many correspondence between filters and classes. Most existing works attempt post-hoc interpretation on a pre-trained model, while neglecting to reduce the entanglement underlying the model. In contrast, we focus on alleviating filter-class entanglement during training. Inspired by cellular differentiation, we propose a novel strategy to train interpretable CNNs by encouraging class-specific filters, among which each filter responds to only one (or few) class. Concretely, we design a learnable sparse Class-Specific Gate (CSG) structure to assign each filter with one (or few) class in a flexible way. The gate allows a filter's activation to pass only when the input samples come from the specific class. Extensive experiments demonstrate the fabulous performance of our method in generating a sparse and highly class-related representation of the input, which leads to stronger interpretability. Moreover, comparing with the standard training strategy, our model displays benefits in applications like object localization and adversarial sample detection. Code link: https://github.com/hyliang96/CSGCNN.
Face recognition has recently made substantial progress and achieved high accuracy on standard benchmarks based on the development of deep convolutional neural networks (CNNs). However, the lack of robustness in deep CNNs to adversarial examples has raised security concerns to enormous face recognition applications. To facilitate a better understanding of the adversarial vulnerability of the existing face recognition models, in this paper we perform comprehensive robustness evaluations, which can be applied as reference for evaluating the robustness of subsequent works on face recognition. We investigate 15 popular face recognition models and evaluate their robustness by using various adversarial attacks as an important surrogate. These evaluations are conducted under diverse adversarial settings, including dodging and impersonation attacks, $\ell_2$ and $\ell_\infty$ attacks, white-box and black-box attacks. We further propose a landmark-guided cutout (LGC) attack method to improve the transferability of adversarial examples for black-box attacks, by considering the special characteristics of face recognition. Based on our evaluations, we draw several important findings, which are crucial for understanding the adversarial robustness and providing insights for future research on face recognition. Code is available at \url{https://github.com/ShawnXYang/Face-Robustness-Benchmark}.
Several machine learning applications involve the optimization of higher-order derivatives (e.g., gradients of gradients) during training, which can be expensive in respect to memory and computation even with automatic differentiation. As a typical example in generative modeling, score matching (SM) involves the optimization of the trace of a Hessian. To improve computing efficiency, we rewrite the SM objective and its variants in terms of directional derivatives, and present a generic strategy to efficiently approximate any-order directional derivative with finite difference (FD). Our approximation only involves function evaluations, which can be executed in parallel, and no gradient computations. Thus, it reduces the total computational cost while also improving numerical stability. We provide two instantiations by reformulating variants of SM objectives into the FD forms. Empirically, we demonstrate that our methods produce results comparable to the gradient-based counterparts while being much more computationally efficient.
Hawkes process provides an effective statistical framework for analyzing the time-dependent interaction of neuronal spiking activities. Although utilized in many real applications, the classical Hawkes process is incapable of modelling inhibitory interactions among neurons. Instead, the nonlinear Hawkes process allows for a more flexible influence pattern with excitatory or inhibitory interactions. In this paper, three sets of auxiliary latent variables (P\'{o}lya-Gamma variables, latent marked Poisson processes and sparsity variables) are augmented to make synapses connection weights in a Gaussian form, which allows for a simple iterative algorithm with analytical updates. As a result, an efficient expectation-maximization (EM) algorithm is derived to obtain the maximum a posteriori (MAP) estimate. We demonstrate the accuracy and efficiency performance of our algorithm on synthetic and real data. For real neural recordings, we show our algorithm can estimate the temporal dynamics of interaction and reveal the interpretable synaptic structure underlying neural spike trains.
Accurate quantification of uncertainty is crucial for real-world applications of machine learning. However, modern deep neural networks still produce unreliable predictive uncertainty, often yielding over-confident predictions. In this paper, we are concerned with getting well-calibrated predictions in regression tasks. We propose the calibrated regression method using the maximum mean discrepancy for distribution level calibration. Theoretically, the calibration error of our method asymptotically converges to zero when the sample size is large enough. Experiments on non-trivial real datasets show that our method can produce well-calibrated and sharp prediction intervals, which outperforms the related state-of-the-art methods.
Granger causality method analyzes the time series causalities without building a complex causality graph. However, the traditional Granger causality method assumes that the causalities lie between time series channels and remain constant, which cannot model the real-world time series data with dynamic causalities along the time series channels. In this paper, we present the dynamic window-level Granger causality method (DWGC) for multi-channel time series data. We build the causality model on the window-level by doing the F-test with the forecasting errors on the sliding windows. We propose the causality indexing trick in our DWGC method to reweight the original time series data. Essentially, the causality indexing is to decrease the auto-correlation and increase the cross-correlation causal effects, which improves the DWGC method. Theoretical analysis and experimental results on two synthetic and one real-world datasets show that the improved DWGC method with causality indexing better detects the window-level causalities.