In a structural health monitoring (SHM) system that uses digital cameras to monitor cracks of structural surfaces, techniques for reliable and effective data compression are essential to ensure a stable and energy efficient crack images transmission in wireless devices, e.g., drones and robots with high definition cameras installed. Compressive sensing (CS) is a signal processing technique that allows accurate recovery of a signal from a sampling rate much smaller than the limitation of the Nyquist sampling theorem. The conventional CS method is based on the principle that, through a regularized optimization, the sparsity property of the original signals in some domain can be exploited to get the exact reconstruction with a high probability. However, the strong assumption of the signals being highly sparse in an invertible space is relatively hard for real crack images. In this paper, we present a new approach of CS that replaces the sparsity regularization with a generative model that is able to effectively capture a low dimension representation of targeted images. We develop a recovery framework for automatic crack segmentation of compressed crack images based on this new CS method and demonstrate the remarkable performance of the method taking advantage of the strong capability of generative models to capture the necessary features required in the crack segmentation task even the backgrounds of the generated images are not well reconstructed. The superior performance of our recovery framework is illustrated by comparing with three existing CS algorithms. Furthermore, we show that our framework is extensible to other common problems in automatic crack segmentation, such as defect recovery from motion blurring and occlusion.

We show how to construct adversarial examples for neural networks with exactly verified robustness against $\ell_{\infty}$-bounded input perturbations by exploiting floating point error. We argue that any exact verification of real-valued neural networks must accurately model the implementation details of any floating point arithmetic used during inference or verification.

Deep probabilistic generative models enable modeling the likelihoods of very high dimensional data. An important application of generative modeling should be the ability to detect out-of-distribution (OOD) samples by setting a threshold on the likelihood. However, a recent study shows that probabilistic generative models can, in some cases, assign higher likelihoods on certain types of OOD samples, making the OOD detection rules based on likelihood threshold problematic. To address this issue, several OOD detection methods have been proposed for deep generative models. In this paper, we make the observation that some of these methods fail when applied to generative models based on Variational Auto-encoders (VAE). As an alternative, we propose Likelihood Regret, an efficient OOD score for VAEs. We benchmark our proposed method over existing approaches, and empirical results suggest that our method obtains the best overall OOD detection performances compared with other OOD method applied on VAE.

Multivariate time series models are poised to be used for decision support in high-stakes applications, such as healthcare. In these contexts, it is important to know which features at which times most influenced a prediction. We demonstrate a general approach for assigning importance to observations in multivariate time series, based on their counterfactual influence on future predictions. Specifically, we define the importance of an observation as the change in the predictive distribution, had the observation not been seen. We integrate over plausible counterfactuals by sampling from the corresponding conditional distributions of generative time series models. We compare our importance metric to gradient-based explanations, attention mechanisms, and other baselines in simulated and clinical ICU data, and show that our approach generates the most precise explanations. Our method is inexpensive, model agnostic, and can be used with arbitrarily complex time series models and predictors.

Label smoothing is commonly used in training deep learning models, wherein one-hot training labels are mixed with uniform label vectors. Empirically, smoothing has been shown to improve both predictive performance and model calibration. In this paper, we study whether label smoothing is also effective as a means of coping with label noise. While label smoothing apparently amplifies this problem --- being equivalent to injecting symmetric noise to the labels --- we show how it relates to a general family of loss-correction techniques from the label noise literature. Building on this connection, we show that label smoothing is competitive with loss-correction under label noise. Further, we show that when distilling models from noisy data, label smoothing of the teacher is beneficial; this is in contrast to recent findings for noise-free problems, and sheds further light on settings where label smoothing is beneficial.

We investigated the effect of different augmentation scenarios on predicting (retro)synthesis of chemical compounds using SMILES representation. We showed that augmentation of not only input sequences but also, importantly, of the target data eliminated the effect of data memorization by neural networks and improved their generalization performance for prediction of new sequences. The Top-5 accuracy was 85.4% for the prediction of the largest fragment (thus identifying principal transformation for classical retro-synthesis) for USPTO-50k test dataset and was achieved by a combination of SMILES augmentation and beam search. The same approach also outperformed best published results for prediction of direct reactions from the USPTO-MIT test set. Our model achieved 90.4% Top-1 and 96.5% Top-5 accuracy for its most challenging mixed set and 97% Top-5 accuracy for the USPTO-MIT separated set. The appearance frequency of the most abundantly generated SMILES was well correlated with the prediction outcome and can be used as a measure of the quality of reaction prediction.

Modern deep networks have millions to billions of parameters, which leads to high memory and energy requirements during training as well as during inference on resource-constrained edge devices. Consequently, pruning techniques have been proposed that remove less significant weights in deep networks, thereby reducing their memory and computational requirements. Pruning is usually performed after training the original network, and is followed by further retraining to compensate for the accuracy loss incurred during pruning. The prune-and-retrain procedure is repeated iteratively until an optimum tradeoff between accuracy and efficiency is reached. However, such iterative retraining adds to the overall training complexity of the network. In this work, we propose a dynamic pruning-while-training procedure, wherein we prune filters of the convolutional layers of a deep network during training itself, thereby precluding the need for separate retraining. We evaluate our dynamic pruning-while-training approach with three different pre-existing pruning strategies, viz. mean activation-based pruning, random pruning, and L1 normalization-based pruning. Our results for VGG-16 trained on CIFAR10 shows that L1 normalization provides the best performance among all the techniques explored in this work with less than 1% drop in accuracy after pruning 80% of the filters compared to the original network. We further evaluated the L1 normalization based pruning mechanism on CIFAR100. Results indicate that pruning while training yields a compressed network with almost no accuracy loss after pruning 50% of the filters compared to the original network and ~5% loss for high pruning rates (>80%). The proposed pruning methodology yields 41% reduction in the number of computations and memory accesses during training for CIFAR10, CIFAR100 and ImageNet compared to training with retraining for 10 epochs .

Deep Learning with noisy labels is a practically challenging problem in weakly-supervised learning. The state-of-the-art approaches "Decoupling" and "Co-teaching+" claim that the "disagreement" strategy is crucial for alleviating the problem of learning with noisy labels. In this paper, we start from a different perspective and propose a robust learning paradigm called JoCoR, which aims to reduce the diversity of two networks during training. Specifically, we first use two networks to make predictions on the same mini-batch data and calculate a joint loss with Co-Regularization for each training example. Then we select small-loss examples to update the parameters of both two networks simultaneously. Trained by the joint loss, these two networks would be more and more similar due to the effect of Co-Regularization. Extensive experimental results on corrupted data from benchmark datasets including MNIST, CIFAR-10, CIFAR-100 and Clothing1M demonstrate that JoCoR is superior to many state-of-the-art approaches for learning with noisy labels.

Measuring the quality of a generated sequence against a set of references is a central problem in many learning frameworks, be it to compute a score, to assign a reward, or to perform discrimination. Despite great advances in model architectures, metrics that scale independently of the number of references are still based on n-gram estimates. We show that the underlying operations, counting words and comparing counts, can be lifted to embedding words and comparing embeddings. An in-depth analysis of BERT embeddings shows empirically that contextual embeddings can be employed to capture the required dependencies while maintaining the necessary scalability through appropriate pruning and smoothing techniques. We cast unconditional generation as a reinforcement learning problem and show that our reward function indeed provides a more effective learning signal than n-gram reward in this challenging setting.

We introduce sentenceMIM, a probabilistic auto-encoder for language modelling, trained with Mutual Information Machine (MIM) learning. Previous attempts to learn variational auto-encoders for language data have had mixed success, with empirical performance well below state-of-the-art auto-regressive models, a key barrier being the occurrence of posterior collapse with VAEs. The recently proposed MIM framework encourages high mutual information between observations and latent variables, and is more robust against posterior collapse. This paper formulates a MIM model for text data, along with a corresponding learning algorithm. We demonstrate excellent perplexity (PPL) results on several datasets, and show that the framework learns a rich latent space, allowing for interpolation between sentences of different lengths with a fixed-dimensional latent representation. We also demonstrate the versatility of sentenceMIM by utilizing a trained model for question-answering, a transfer learning task, without fine-tuning. To the best of our knowledge, this is the first latent variable model (LVM) for text modelling that achieves competitive performance with non-LVM models.