Greedy Policy Search: A Simple Baseline for Learnable Test-Time Augmentation
Test-time data augmentation—averaging the predictions of a machine learning model across multiple augmented samples of data—is a widely used technique that improves the predictive performance. While many advanced learnable data augmentation techniques have emerged in recent years, they are focused on the training phase. Such techniques are not necessarily optimal for test-time augmentation and can be outperformed by a policy consisting of simple crops and flips. The primary goal of this paper is to demonstrate that test-time augmentation policies can be successfully learned too. We introduce greedy policy search (GPS), a simple but high-performing method for learning a policy of test-time augmentation. We demonstrate that augmentation policies learned with GPS achieve superior predictive performance on image classification problems, provide better in-domain uncertainty estimation, and improve the robustness to domain shift.
The series “Lecture Notes in Networks and Systems” publishes the latest develop- ments in Networks and Systems—quickly, informally and with high quality. Original research reported in proceedings and post-proceedings represents the core of LNNS.
Volumes published in LNNS embrace all aspects and subfields of, as well as new challenges in, Networks and Systems.
The series contains proceedings and edited volumes in systems and networks, spanning the areas of Cyber-Physical Systems, Autonomous Systems, Sensor Networks, Control Systems, Energy Systems, Automotive Systems, Biological Systems, Vehicular Networking and Connected Vehicles, Aerospace Systems, Automation, Manufacturing, Smart Grids, Nonlinear Systems, Power Systems, Robotics, Social Systems, Economic Systems and other. Of particular value to both the contributors and the readership are the short publication timeframe and the world-wide distribution and exposure which enable both a wide and rapid dissemination of research output.
The series covers the theory, applications, and perspectives on the state of the art and future developments relevant to systems and networks, decision making, control, complex processes and related areas, as embedded in the fields of interdisciplinary and applied sciences, engineering, computer science, physics, economics, social, and life sciences, as well as the paradigms and methodologies behind them.
This article describes modern methods of data processing regarding the task of assessing activities of transportation employees. The main purpose was to find dependencies in data and construct an algorithm for predicting the probability of transport safety violation by employee. The research was conducted for locomotive drivers. The following algorithms were used: neural networks, gradient boosting over decision trees and random forest. Based on the obtained results and drawn conclusions one can think of the perspective for the elaboration and introduction this work for practical use in railway industry, e.g. in “Russian Railways”.
We compared the ability of various empirical methods to reproduce public credit ratings (PCRs) of industrial companies (ICs) from BRICS countries using publicly available information. This task is important for researchers and practitioners because many of BRICS’ ICs lack PCRs from reputable rating agencies such as Moody’s, Standard and Poor’s, and Fitch. This paper aimed at filling the gap in the existing research as insufficient efforts were focused on prediction of PCRs of ICs from the entire BRICS IC community. The modeled variables are credit ratings (CRs) of 208 BRICS’ ICs assigned by Moody’s at the year-end from 2006 to 2016. The sample included 1217 observations. Financial explanatory variables included companies’ revenue, operating profitability, interest coverage ratio, debt/book capitalization, and cash flow debt coverage. Non-financial explanatory variables included dummies for home region, industry, affiliation with the state, and a set of macroeconomic data of IC’s home countries. The set of statistical methods included linear discriminant analysis (LDA), ordered logit regression (OLR), support vector machine (SVM), artificial neural network (ANN), and random forest (RF). The resulting models were checked for in-sample and out-of-sample predictive fit. Our findings revealed that among considered methods of artificial intelligence models (AI), SVM, ANN, and RF outperformed LDA and OLR by predictive power. On testing sample, AI gave on average 55% of precise results and up to 99% with an error within one rating grade; RF demonstrated the best outcome (58% and 100%). Conversely, LDA and OLR on average gave only 37% of precise results and up to 70% with an error within one grade. LDA and OLR also gave higher share of Type I errors (overestimation of ratings) than that of AI. Therefore, AI should have higher practical application than DA and OLR for predicting the ratings of BRICS ICs
The focus of the paper is on the study of the emergence of the market for artificial intelligence technologies in Russia based on both expert poll and survey of CEOs of the Russian industrial enterprises. It includes two parts. The first contains the methodology of the research, a description of the market agents, and the features of the product. In the second part, the authors analyze the interactions of the agents and the role of the state in the regulation of the market. This emerging market combines the features of the markets for software products and consulting services, which offer solutions, i.e. unique personalized products tailored to the needs and conditions of specific companies. In spite of fast growth, the development of the market faced significant obstacles, which can slow it down in the future.
The purpose of the paper is to determine the perspectives of diversification of educational services in the conditions of industry 4.0 on the basis of artificial intelligence (AI) training, determine the consequences of this process for academic and teaching staff and to develop recommendations for its practical implementation.
In November 2014, Team DESCARTES led by Newton Lee and sponsored by the Institute for Education, Research, and Scholarships (IFERS) was among one of the 104 teams registered with the Defense Advanced Research Projects Agency (DARPA) for the first-ever Cyber Grand Challenge (CGC). Only 28 teams, including Team DESCARTES, made it through two DARPA-sponsored dry runs and into the CGC Qualifying Event in June 2015. We proposed a system—Distributed Expert Systems for Cyber Analysis, Reasoning, Testing, Evaluation, and Security (DESCARTES)—that would be a fully autonomous cyber defense system that is capable of autonomous analysis, autonomous patching, autonomous vulnerability scanning, autonomous service resiliency, and autonomous network defense.
Logical frameworks allow the specification of deductive systems using the same logical machinery. Linear logical frameworks have been successfully used for the specification of a number of computational, logics and proof systems. Its success relies on the fact that formulas can be distinguished as linear, which behave intuitively as resources, and unbounded, which behave intuitionistically. Commutative subexponentials enhance the expressiveness of linear logic frameworks by allowing the distinction of multiple contexts. These contexts may behave as multisets of formulas or sets of formulas. Motivated by applications in distributed systems and in type-logical grammar, we propose a linear logical framework containing both commutative and non-commutative subexponentials. Non-commutative subexponentials can be used to specify contexts which behave as lists, not multisets, of formulas. In addition, motivated by our applications in type-logical grammar, where the weakenening rule is disallowed, we investigate the proof theory of formulas that can only contract, but not weaken. In fact, our contraction is non-local. We demonstrate that under some conditions such formulas may be treated as unbounded formulas, which behave intuitionistically.