The key distinguishing property of a Bayesian approach is marginalization instead of optimization, not the prior, or Bayes rule. Bayesian inference is especially compelling for deep neural networks. (1) Neural networks are typically underspecified by the data, and can represent many different but high performing models corresponding to different settings of parameters, which is exactly when marginalization will make the biggest difference for both calibration and accuracy. (2) Deep ensembles have been mistaken as competing approaches to Bayesian methods, but can be seen as approximate Bayesian marginalization. (3) The structure of neural networks gives rise to a structured prior in function space, which reflects the inductive biases of neural networks that help them generalize. (4) The observed correlation between parameters in flat regions of the loss and a diversity of solutions that provide good generalization is further conducive to Bayesian marginalization, as flat regions occupy a large volume in a high dimensional space, and each different solution will make a good contribution to a Bayesian model average. (5) Recent practical advances for Bayesian deep learning provide improvements in accuracy and calibration compared to standard training, while retaining scalability.

--Capsule networks are a recently developed class of neural networks that potentially address some of the deficiencies with traditional convolutional neural networks. By replacing the standard scalar activations with vectors, and by connecting the artificial neurons in a new way, capsule networks aim to be the next great development for computer vision applications. However, in order to determine whether these networks truly operate differently than traditional networks, one must look at the differences in the capsule features. T o this end, we perform several analyses with the purpose of elucidating capsule features and determining whether they perform as described in the initial publication. First, we perform a deep visualization analysis to visually compare capsule features and convolutional neural network features. Then, we look at the ability for capsule features to encode information across the vector components and address what changes in the capsule architecture provides the most benefit. Finally, we look at how well the capsule features are able to encode instantiation parameters of class objects via visual transformations. ONVOLUTIONAL neural networks (CNNs) have long been the tools of choice when tackling computer vision problems. The spatial localization of CNN features is greatly beneficial when the networks are applied to images and videos; however, these networks also have their shortcomings. The kernels in a convolutional layer must learn to identify the presence of all relevant features in the input. Thus, transformations such as rotations and occlusion can be detrimental when the training dataset is not properly augmented. Even still, the burden of learning visual features in addition to all possible modifications of these features can be immense for a traditional CNN. Recently, a novel class of neural networks was proposed in [1] that employs the concept of a "capsule".

Obtaining labels for medical (image) data requires scarce and expensive experts. Moreover, due to ambiguous symptoms, single images rarely suffice to correctly diagnose a medical condition. Instead, it often requires to take additional background information such as the patient's medical history or test results into account. Hence, instead of focusing on uninterpretable black-box systems delivering an uncertain final diagnosis in an end-to-end-fashion, we investigate how unsupervised methods trained on images without anomalies can be used to assist doctors in evaluating X-ray images of hands. Our method increases the efficiency of making a diagnosis and reduces the risk of missing important regions. Therefore, we adopt state-of-the-art approaches for unsupervised learning to detect anomalies and show how the outputs of these methods can be explained. To reduce the effect of noise, which often can be mistaken for an anomaly, we introduce a powerful preprocessing pipeline. We provide an extensive evaluation of different approaches and demonstrate empirically that even without labels it is possible to achieve satisfying results on a real-world dataset of X-ray images of hands. We also evaluate the importance of preprocessing and one of our main findings is that without it, most of our approaches perform not better than random. To foster reproducibility and accelerate research we make our code publicly available at https://github.com/Valentyn1997/xray

Not all data in a typical training set help with generalization; some samples can be overly ambiguous or outrightly mislabeled. This paper introduces a new method to identify such samples and mitigate their impact when training neural networks. At the heart of our algorithm is the Area Under the Margin (AUM) statistic, which exploits differences in the training dynamics of clean and mislabeled samples. A simple procedure - adding an extra class populated with purposefully mislabeled indicator samples - learns a threshold that isolates mislabeled data based on this metric. This approach consistently improves upon prior work on synthetic and real-world datasets. On the WebVision50 classification task our method removes 17% of training data, yielding a 2.6% (absolute) improvement in test error. On CIFAR100 removing 13% of the data leads to a 1.2% drop in error.

Network representation learning (NRL) is a powerful technique for learning low-dimensional vector representation of high-dimensional and sparse graphs. Most studies explore the structure and metadata associated with the graph using random walks and employ an unsupervised or semi-supervised learning schemes. Learning in these methods is context-free, because only a single representation per node is learned. Recently studies have argued on the sufficiency of a single representation and proposed a context-sensitive approach that proved to be highly effective in applications such as link prediction and ranking. However, most of these methods rely on additional textual features that require RNNs or CNNs to capture high-level features or rely on a community detection algorithm to identify multiple contexts of a node. In this study, without requiring additional features nor a community detection algorithm, we propose a novel context-sensitive algorithm called GAP that learns to attend on different parts of a node's neighborhood using attentive pooling networks. We show the efficacy of GAP using three real-world datasets on link prediction and node clustering tasks and compare it against 10 popular and state-of-the-art (SOTA) baselines. GAP consistently outperforms them and achieves up to ~9% and ~20% gain over the best performing methods on link prediction and clustering tasks, respectively.

Humans are skilled at learning adaptively and efficiently throughout their lives, but learning tasks incrementally causes artificial neural networks to overwrite relevant information learned about older tasks, resulting in 'Catastrophic Forgetting'. Efforts to overcome this phenomenon suffer from poor utilization of resources in many ways, such as through the need to save older data or parametric importance scores, or to grow the network architecture. We propose an algorithm that enables a network to learn continually and efficiently by partitioning the representational space into a Core space, that contains the condensed information from previously learned tasks, and a Residual space, which is akin to a scratch space for learning the current task. The information in the Residual space is then compressed using Principal Component Analysis and added to the Core space, freeing up parameters for the next task. We evaluate our algorithm on P-MNIST, CIFAR-10 and CIFAR-100 datasets. We achieve comparable accuracy to state-of-the-art methods while overcoming the problem of catastrophic forgetting completely. Additionally, we get up to 4.5x improvement in energy efficiency during inference due to the structured nature of the resulting architecture.

The human visual system is remarkably robust against a wide range of naturally occurring variations and corruptions like rain or snow. In contrast, the performance of modern image recognition models strongly degrades when evaluated on previously unseen corruptions. Here, we demonstrate that a simple but properly tuned training with additive Gaussian and Speckle noise generalizes surprisingly well to unseen corruptions, easily reaching the previous state of the art on the corruption benchmark ImageNet-C (with ResNet50) and on MNIST-C. We build on top of these strong baseline results and show that an adversarial training of the recognition model against uncorrelated worst-case noise distributions leads to an additional increase in performance. This regularization can be combined with previously proposed defense methods for further improvement.

As deep neural networks are increasingly being deployed in practice, their efficiency has become an important issue. While there are compression techniques for reducing the network's size, energy consumption and computational requirement, they only demonstrate empirically that there is no loss of accuracy, but lack formal guarantees of the compressed network, e.g., in the presence of adversarial examples. Existing verification techniques such as Reluplex, ReluVal, and DeepPoly provide formal guarantees, but they are designed for analyzing a single network instead of the relationship between two networks. To fill the gap, we develop a new method for differential verification of two closely related networks. Our method consists of a fast but approximate forward interval analysis pass followed by a backward pass that iteratively refines the approximation until the desired property is verified. We have two main innovations. During the forward pass, we exploit structural and behavioral similarities of the two networks to more accurately bound the difference between the output neurons of the two networks. Then in the backward pass, we leverage the gradient differences to more accurately compute the most beneficial refinement. Our experiments show that, compared to state-of-the-art verification tools, our method can achieve orders-of-magnitude speedup and prove many more properties than existing tools.

We access the internalized understanding of trained, deep neural networks to perform Bayesian reasoning on complex tasks. Independently trained networks are arranged to jointly answer questions outside their original scope, which are formulated in terms of a Bayesian inference problem. We solve this approximately with variational inference, which provides uncertainty on the outcomes. We demonstrate how following tasks can be approached this way: Combining independently trained networks to sample from a conditional generator, solving riddles involving multiple constraints simultaneously, and combine deep-learned knowledge with conventional noisy measurements in the context of high-resolution images of human faces.

We analyse perception and memory using mathematical models for knowledge graphs and tensors to gain insights in the corresponding functionalities of the human mind. Our discussion is based on the concept of propositional sentences consisting of \textit{subject-predicate-object} (SPO) triples for expressing elementary facts. SPO sentences are the basis for most natural languages but might also be important for explicit perception and declarative memories, as well as intra-brain communication and the ability to argue and reason. A set of SPO sentences can be described as a knowledge graph, which can be transformed into an adjacency tensor. We introduce tensor models, where concepts have dual representations as indices and associated embeddings, two constructs we believe are essential for the understanding of implicit and explicit perception and memory in the brain. We argue that a biological realization of perception and memory imposes constraints on information processing. In particular, we propose that explicit perception and declarative memories require a semantic decoder, which, in a simple realization, is based on four layers: First, a sensory memory layer, as a buffer for sensory input, second, an index layer representing concepts, third, a memoryless representation layer for the broadcasting of information and fourth, a working memory layer as a processing center and data buffer. In a Bayesian brain interpretation, semantic memory defines the prior for triple statements. We propose that, in evolution and during development, semantic memory, episodic memory and natural language evolved as emergent properties in the agents' process to gain deeper understanding of sensory information. We present a concrete model realization and validate some aspects of our proposed model on benchmark data where we demonstrate state-of-the-art performance.