One of the most important optimization tasks in the industry at the current time is the object packing problem. Although several methods have been developed for the purpose of solving it, they are usually only able to optimize placement locally and as such are heavily dependent on the choice of the initial setting -- hence the need for trying out multiple possible starting points, which impacts algorithm running time. In this paper we present a neural network-based model which provides sensible starting points in a linear time.

We investigate performance of a gradient descent optimization (GR) applied to the traffic signal setting problem and compare it to genetic algorithms. We used neural networks as metamodels evaluating quality of signal settings and discovered that both optimization methods produce similar results, e.g., in both cases the accuracy of neural networks close to local optima depends on an activation function (e.g., TANH activation makes optimization process converge to different minima than ReLU activation).

Regularizing the gradient norm of the output of a neural network is a powerful technique, rediscovered several times. This paper presents evidence that gradient regularization can consistently improve classification accuracy on vision tasks, using modern deep neural networks, especially when the amount of training data is small. We introduce our regularizers as members of a broader class of Jacobian-based regularizers. We demonstrate empirically on real and synthetic data that the learning process leads to gradients controlled beyond the training points, and results in solutions that generalize well.

In this paper, we propose a simple, fast and easy to implement algorithm LOSSGRAD (locally optimal step-size in gradient descent), which automatically modifies the step-size in gradient descent during neural networks training. Given a function f, a point x, and the gradient ▽_xf of f, we aim to find the step-size h which is (locally) optimal, i.e. satisfies:

h = arg min f(x - t▽_xf).
            t≥0

Making use of quadratic approximation, we show that the algorithm satisfies the above assumption. We experimentally show that our method is insensitive to the choice of initial learning rate while achieving results comparable to other methods.

We apply the optimal transport distance to construct two goodness of fit tests for (univariate) normality. The derived statistics are then compared with those used by the Shapiro-Wilk, the Anderson-Darling and the Cramer-von Mises tests. In particular, we preform Monte Carlo experiments, involving computations of the test power against some selected alternatives and wide range of sample sizes, which show efficiency of the obtained test procedures.

In this paper we discuss a class of  AutoEncoder based generative models based on one dimensional sliced approach. The idea is based on the reduction of the discrimination between samples to one-dimensional case.

Our experiments show that methods can be divided into two groups. First consists of methods which are a modification of standard normality tests, while the second is based on classical distances between samples.

It turns out that both groups are correct generative models, but the second one gives a slightly faster decrease rate of Frechet Inception Distance (FID).

We present a novel modification of context encoder loss function, which results in more accurate and plausible inpainting. For this purpose, we introduce gradient attention loss component of loss function, to suppress the common problem of inconsistency in shapes and edges between the inpainted region and its context. To this end, the mean absolute error is computed not only for the input and output images, but also for their derivatives. Therefore, model concentrates on areas with larger gradient, which are crucial for accurate reconstruction. The positive effects on inpainting results are observed both for fully-connected and fully-convolutional models tested on MNIST and CelebA datasets.

The article introduces a new set of Polish word embeddings, built using KGR10 corpus, which contains more than 4 billion words. These embeddings are evaluated in the problem of recognition of temporal expressions (timexes) for the Polish language. We described the process of KGR10 corpus creation and a new approach to the recognition problem using Bidirectional Long-Short Term Memory (BiLSTM) network with additional CRF layer, where specific embeddings are essential. We presented experiments and conclusions drawn from them.

Infinite length of trajectories is an almost universal assumption in the theoretical foundations of reinforcement learning. In practice learning occurs on finite trajectories. In this paper we examine a specific result of this disparity, namely a strong bias of the time-bounded Every-visit Monte Carlo value estimator. This manifests as a vastly different learning dynamic for algorithms that use value predictors, including encouraging or discouraging exploration. We investigate these claims theoretically for a one dimensional random walk, and empirically on a number of simple environments. We use GAE as an algorithm involving a value predictor and evolution strategies as a reference point.

In this paper we present a method with closed analytic formula of stitching aligned images.

It is obtained by choosing a statistically optimal global color change of one part of image.  This approach, due to its numerical efficiency, is especially well-suited for merging big amount of satellite images into a single map.

Moreover, we present solution of a general problem, how to find an optimal shift by v of data Y with respect to v from V, so that the dataset  X, Y+v is maximally statistically consistent. We show that the solution is given in a closed analytic form.

Natural language inference (NLI) is a central problem in natural language processing (NLP) of predicting the logical relationship between a pair of sentences. Lexical knowledge, which represents relations between words, is often important for solving NLI problems. This knowledge can be accessed by using an external knowledge base (KB), but this is limited to when such a resource is accessible. Instead of using a KB, we propose a simple architectural change for attention based models. We show that by adding a skip connection from the input to the attention layer we can utilize better the lexical knowledge already present in the pretrained word embeddings. Finally, we demonstrate that our strategy allows to use an external source of knowledge in a straightforward manner by incorporating a second word embedding space in the model.

Neural networks are prone to adversarial attacks. In general, such attacks deteriorate the quality of the input by either slightly modifying most of its pixels, or by occluding it with a patch. In this paper, we propose a method that keeps the image unchanged and only adds an adversarial framing on the border of the image. We show empirically that our method is able to successfully attack state-of-the-art methods on both image and video classification problems. Notably, the proposed method results in a universal attack which is very fast at test time.

Credibility is an important part of any quality scoring method. Scoring quality of a piece of information in regard to other messages gives an additional and crucial point of view. This quality dimension is especially useful when evaluating information produced by sensors. Often grouped in a network, a message from one sensor can be evaluated using other sources in this network which can highlight problematic messages leading to improved decision making or optimising maintenance operations.

This paper considers the problem of defining credibility on sensor data: including definition of confirmation and invalidation for a piece of information in the case of the more challenging event-type messages. The proposed method aims to define a network of correlated sensors from which the relevant messages are extracted and then aggregated.

Different propositions for the final aggregation step of confirming and invalidating messages leads to the definition of a flexible framework that can be adapted to different scenarios.

An example is presented based on a real dataset that includes sensors from railway domain.