Искусственные нейронные сети как особый тип distant reading
The article deals with the generation of poetic texts with artificial neural networks. The author gives a brief history of the method. The article describes some important properties of the training sample. For example the sample needs to be large enough. The article gives some examples of Russian poetic texts, generated by a neural network. The texts generated by the model trained on Russian hexameters, on poems of a modern poet Natalia Azarova and on texts of classic Russian bard Vladimir Vysotsky. The analysis showed that the neural network reproduces the style and metrical features of the original sample. The style of the lyrical texts reproduced better than any other type of text. A neural network is practically unable to reproduce the features of narrative works. In the cultural and intellectual context, the texts of the neural network can be understood as deconstruction (Derrida) and reassembling (Latour).
This volume constitutes the refereed proceedings of the 4th International Conference on Digital Transformation and Global Society, DTGS 2019, held in St. Petersburg, Russia, in June 2019.
Reservoir Computing (RC) is taking attention of neural networks structures developers because of machine learning algorithms are simple at the high level of generalization of the models. The approaches are numerous. RC can be applied to different architectures including recurrent neural networks with irregular connections that are called Echo State Networks (ESN). However, the existence of successful examples of chaotic sequences predictions does not provide successful method of multiple attribute objects classification.
In this paper the binary ESN classifiers are researched. We show that the reason of low precision of classification is the existence of unbalanced classes. Then the method to solve the problem is proposed. It is possible to use randomizing algorithm of learning data set balancing and method of data temporalization. The resulting errors matrixes have pretty good numbers. The proposed method is illustrated by the usage on synthetic data set. The features of ESN classifier are demonstrated in the case of rare events detection such as transaction attributes fraud detection.
This research examines the problems of automatic scientific articles classification according to Universal Decimal Classifier. To reveal the structure of the train data its visualization was obtained using the recursive feature elimination algorithm. Further; the study provides a comparison of TF-IDF and Weirdness – two statistic-based metrics of keyword significance. The most efficient classification methods are explained: cosine similarity method, naïve Bayesian classifier and artificial neural network. This research explores the most effective for text categorization structure of the multi-layer perceptron and derives appropriate conclusions.
The paper briefly introduces the history of cognitive psychology from its emergence in the 1950s until the present. The unique contribution of cognitive psychology to psychological science is discussed. The main lines of cognitive psychology criticism and self-criticism are outlined: they include the single representational format in the information processing system, the limited resources of this system, and the degree of similarity in information processing between living and artificial systems. A number of state of the art research areas have emerged as a response to these criticisms: among them are embodied cognition, situated cognition, social and distributed cognition, emotional cognition, and many others. Possible scenarios of the further development of cognitive psychology and cognitive science are analyzed.
Popularity of social networks makes them an attractive field for analysis of users’ behavior, for example, based on the intention analysis of their posts and comments. In the linguistic theory only 25 types of intentions exist and can be joined in 5 supergroups. We use the dataset that contains directed oriented graphs which nodes store information about the author intention, text of the post in the social network “Vkontakte” etc. Each graph is split in a linked list of nodes (a sequence, 13156 sequences in our dataset) from root to each leaf so that the intention prediction becomes the sequence prediction. We have analyzed traditional and neuronet approaches that address this task and proposed to solve it with the original modifications of CNN and RNN architectures. It was decided to translate all posts to the embeddings which are then used as inputs for our neural network. According to the benchmarking experiments, we have identified that the proposed RNN architecture outperforms other alternatives. Also, predicting supergroups is done more accurately. Finally, we found out, that the context in the dialogs is lost quickly that allows to decrease the algorithm context size while keeping accuracy at the appropriate level.
A task of maximizing deep learning neural networks performance is a challenging and actual goal of modern hardware and software development. Regardless the huge variety of optimization techniques and emerging dedicated hardware platforms, the process of tuning the performance of the neural network is hard. It requires configuring dozens of hyper parameters of optimization algorithms, selecting appropriate metrics, benchmarking the intermediate solutions to choose the best method, platform etc. Moreover, it is required to setup the hardware for the specific inference target. This paper introduces a sophisticated software solution (Deep Learning Workbench) that provides interactive user interface, simplified process of 8-bit quantization, speeding up convolutional operations using the Winograds minimal filtering algorithms, measuring accuracy of the resulting model. The proposed software is built over the open source OpenVINO framework and supports huge range of modern deep learning models.
ICCV is the premier international computer vision event comprising the main conference and several co-located workshops and tutorials. With its high quality and low cost, it provides an exceptional value for students, academics and industry researchers.
2018 proceedings of European Symposium on Artificial Neural Networks, Computational Intelligence and Machine Learning (ESANN 2018),Bruges (Belgium), 25-27 April 2018 discusses the problems of deep learning and image processing, regression and recommendation systems, Extreme Minimal Learning Machine, Neural networks, Facial emotion recognition.
The paper is focused on the study of reaction of italian literature critics on the publication of the Boris Pasternak's novel "Doctor Jivago". The analysys of the book ""Doctor Jivago", Pasternak, 1958, Italy" (published in Russian language in "Reka vremen", 2012, in Moscow) is given. The papers of italian writers, critics and historians of literature, who reacted immediately upon the publication of the novel (A. Moravia, I. Calvino, F.Fortini, C. Cassola, C. Salinari ecc.) are studied and analised.
In the article the patterns of the realization of emotional utterances in dialogic and monologic speech are described. The author pays special attention to the characteristic features of the speech of a speaker feeling psychic tension and to the compositional-pragmatic peculiarities of dialogic and monologic text.