Deep neural networks (DNNs) have surpassed human-level accuracy in various learning tasks. However, unlike humans who have a natural cognitive intuition for probabilities, DNNs cannot express their uncertainty in the output decisions. This limits the deployment of DNNs in mission-critical domains, such as warfighter decision-making or medical diagnosis. Bayesian inference provides a principled approach to reason about model's uncertainty by estimating the posterior distribution of the unknown parameters. The challenge in DNNs remains the multi-layer stages of non-linearities, which make the propagation of high-dimensional distributions mathematically intractable. This paper establishes the theoretical and algorithmic foundations of uncertainty or belief propagation by developing new deep learning models named PremiUm-CNNs (Propagating Uncertainty in Convolutional Neural Networks). We introduce a tensor normal distribution as a prior over convolutional kernels and estimate the variational posterior by maximizing the evidence lower bound (ELBO). We start by deriving the first-order mean-covariance propagation framework. Later, we develop a framework based on the unscented transformation (correct at least up to the second-order) that propagates sigma points of the variational distribution through layers of a CNN. The propagated covariance of the predictive distribution captures uncertainty in the output decision. Comprehensive experiments conducted on diverse benchmark datasets demonstrate: 1) superior robustness against noise and adversarial attacks, 2) self-assessment through predictive uncertainty that increases quickly with increasing levels of noise or attacks, and 3) an ability to detect a targeted attack from ambient noise.

Subsurface fires and smoldering events at landfills can present serious health hazards and threats to the environment. These fires are much more costly and difficult to extinguish than open fires at the landfill surface. The initiation of a subsurface fire may go unnoticed for a long period of time and undetected fires may spread over a large area. Unfortunately, not all landfill operators keep or publish heat elevation data and many landfills are not equipped with a landfill gas extraction system to control subsurface temperatures generated from the chemical reactions within. The timely and cost-effective identification of subsurface fires is an important and pressing issue. In this work, we describe a method for using satellite thermal infrared imagery at a moderate spatial resolution to identify the locations of subsurface fires and monitor their migration within landfills. The focus of this study was the Bridgeton Sanitary Landfill in Bridgeton, MO, USA where a subsurface fire was first identified in 2010 and continues to burn today. Observations from Landsat satellites over the last seventeen years were examined for surface temperature anomalies (or hot spots) that may be associated with subsurface fires. The results showed that the locations of hot spots identified in satellite imagery match the known locations of the subsurface fires. Changes in the hot-spot locations with time, as determined by in situ measurements, correspond to the spreading routes of the subsurface fires. These results indicate that the proposed approach based on satellite observations can be used as a tool for the identification of landfill subsurface fires by landfill owners/operators to monitor landfills and minimize the expenses associated with extinguishing landfill fires.

Image restoration and recognition are important computer vision tasks representing an inherent part of autonomous systems. These two tasks are often implemented in a sequential manner, in which the restoration process is followed by a recognition. In contrast, this paper proposes a joint framework that simultaneously performs both tasks within a shared deep neural network architecture. This joint framework integrates the restoration and recognition tasks by incorporating: (i) common layers, (ii) restoration layers and (iii) classification layers. The total loss function combines the restoration and classification losses. The proposed joint framework, based on capsules, provides an efficient solution that can cope with challenges due to noise, image rotations and occlusions. The developed framework has been validated and evaluated on a public vehicle logo dataset under various degradation conditions, including Gaussian noise, rotation and occlusion. The results show that the joint framework improves the accuracy compared with the single task networks.

The massive time-series production through the Internet of Things and digital healthcare requires novel data modeling and prediction. Recurrent neural networks (RNNs) are extensively used for analyzing time-series data. However, these models are unable to assess prediction uncertainty, which is particularly critical in heterogeneous and noisy environments. Bayesian inference allows reasoning about predictive uncertainty by estimating the posterior distribution of the parameters. The challenge remains in propagating the high-dimensional distribution through the sequential, non-linear layers of RNNs, resulting in mode collapse leading to erroneous uncertainty estimation and exacerbating the gradient explosion problem. This paper proposes a TRustworthy Uncertainty propagation for Sequential Time-series analysis (TRUST) in RNNs by introducing a Gaussian prior over network parameters and estimating the first two moments of the Gaussian variational distribution using the evidence lower bound. We propagate the variational moments through the sequential, non-linear layers of RNNs using the first-order Taylor approximation. The propagated covariance of the predictive distribution captures uncertainty in the output decision. The extensive experiments using ECG5000 and PeMS-SF classification and weather and power consumption prediction tasks demonstrate 1) significant robustness of TRUST-RNNs against noise and adversarial attacks and 2) self-assessment through the uncertainty that increases significantly with increasing noise.

Artificial intelligence and neuroscience have a long and intertwined history. Advancements in neuroscience research have significantly influenced the development of artificial intelligence systems that have the potential to retain knowledge akin to humans. Building upon foundational insights from neuroscience and existing research in adversarial and continual learning fields, we introduce a novel framework that comprises two key concepts: feature distillation and re-consolidation. The framework distills continual learning (CL) robust features and rehearses them while learning the next task, aiming to replicate the mammalian brain’s process of consolidating memories through rehearsing the distilled version of the waking experiences. Furthermore, the proposed framework emulates the mammalian brain’s mechanism of memory re-consolidation, where novel experiences influence the assimilation of previous experiences via feature re-consolidation. This process incorporates the new understanding of the CL model after learning the current task into the CL-robust samples of the previous task(s) to mitigate catastrophic forgetting. The proposed framework, called Robust Rehearsal, circumvents the limitations of existing CL frameworks that rely on the availability of pre-trained Oracle CL models to pre-distill CL-robustified datasets for training subsequent CL models. We conducted extensive experiments on three datasets, CIFAR10, CIFAR100, and real-world helicopter attitude datasets, demonstrating that CL models trained using Robust Rehearsal outperform their counterparts’ baseline methods. In addition, we conducted a series of experiments to assess the impact of changing memory sizes and the number of tasks, demonstrating that the baseline methods employing robust rehearsal outperform other methods trained without robust rehearsal. Lastly, to shed light on the existence of diverse features, we explore the effects of various optimization training objectives within the realms of joint, continual, and adversarial learning on feature learning in deep neural networks. Our findings indicate that the optimization objective dictates feature learning, which plays a vital role in model performance. Such observation further emphasizes the importance of rehearsing the CL-robust samples in alleviating catastrophic forgetting. In light of our experiments, closely following neuroscience insights can contribute to developing CL approaches to mitigate the long-standing challenge of catastrophic forgetting.