When building a unified vision system or gradually adding new capabilities to a system, the usual assumption is that training data for all tasks is always available. However, as the number of tasks grows, storing and retraining on such data becomes infeasible. A new problem arises where we add new capabilities to a Convolutional Neural Network (CNN), but the training data for its existing capabilities are unavailable. We propose our Learning without Forgetting method, which uses only new task data to train the network while preserving the original capabilities. Our method performs favorably compared to commonly used feature extraction and fine-tuning adaption techniques and performs similarly to multitask learning that uses original task data we assume unavailable. A more surprising observation is that Learning without Forgetting may be able to replace fine-tuning with similar old and new task datasets for improved new task performance.

Learning without Forgetting

The objective of change-point detection is to discover abrupt property changes lying behind time-series data. In this paper, we present a novel statistical change-point detection algorithm based on non-parametric divergence estimation between time-series samples from two retrospective segments. Our method uses the relative Pearson divergence as a divergence measure, and it is accurately and efficiently estimated by a method of direct density-ratio estimation. Through experiments on artificial and real-world datasets including human-activity sensing, speech, and Twitter messages, we demonstrate the usefulness of the proposed method.

Change-Point Detection in Time-Series Data by Relative Density-Ratio Estimation (情報論的学習理論と機械学習)

Anonymization is a practical solution for preserving user’s privacy in data publishing. Data owners such as hospitals, banks, social network (SN) service providers, and insurance companies anonymize their user’s data before publishing it to protect the privacy of users whereas anonymous data remains useful for legitimate information consumers. Many anonymization models, algorithms, frameworks, and prototypes have been proposed/developed for privacy preserving data publishing (PPDP). These models/algorithms anonymize users’ data which is mainly in the form of tables or graphs depending upon the data owners. It is of paramount importance to provide good perspectives of the whole information privacy area involving both tabular and SN data, and recent anonymization researches. In this paper, we presents a comprehensive survey about SN (i.e., graphs) and relational (i.e., tabular) data anonymization techniques used in the PPDP. We systematically categorize the existing anonymization techniques into relational and structural anonymization, and present an up to date thorough review on existing anonymization techniques and metrics used for their evaluation. Our aim is to provide deeper insights about the PPDP problem involving both graphs and tabular data, possible attacks that can be launched on the sanitized published data, different actors involved in the anonymization scenario, and major differences in amount of private information contained in graphs and relational data, respectively. We present various representative anonymization methods that have been proposed to solve privacy problems in application-specific scenarios of the SNs. Furthermore, we highlight the user’s re-identification methods used by malevolent adversaries to re-identify people uniquely from the privacy preserved published data. Additionally, we discuss the challenges of anonymizing both graphs and tabular data, and elaborate promising research directions. To the best of our knowledge, this is the first work to systematically cover recent PPDP techniques involving both SN and relational data, and it provides a solid foundation for future studies in the PPDP field.

Anonymization Techniques for Privacy Preserving Data Publishing: A Comprehensive Survey

Pluvial flooding can have devastating effects, both in terms of loss of life and damage. Predicting pluvial floods is difficult and many cities do not have a hydrodynamic model or an early warning system in place. Citizen science and crowdsourcing have the potential for contributing to early warning systems and can also provide data for validating flood forecasting models. Although there are increasing applications of citizen science and crowdsourcing in fluvial hydrology, less is known about activities related to pluvial flooding. Hence the aim of this paper is to review current activities in citizen science and crowdsourcing with respect to applications of pluvial flooding. Based on a search in Scopus, the papers were first filtered for relevant content and then classified into four main themes. The first two themes were divided into (i) applications relevant during a flood event, which includes automated street flooding detection using crowdsourced photographs and sensors, analysis of social media, and online and mobile applications for flood reporting; and (ii) applications related to post-flood events. The use of citizen science and crowdsourcing for model development and validation is the third theme while the development of integrated systems is theme four. All four main areas of research have the potential to contribute to early warning systems and build community resilience. Moreover, developments in one will benefit others, e.g., further developments in flood reporting applications and automated flood detection systems will yield data useful for model validation.

/pdf/a-review-of-citizen-science-and-crowdsourcing-in-297avsed98.pdf

A Review of Citizen Science and Crowdsourcing in Applications of Pluvial Flooding

Condition monitoring of industrial equipment, combined with machine learning algorithms, may significantly improve maintenance activities on modern cyber-physical production systems. However, data of proper quality and of adequate quantity, modeling both good operational conditions as well as abnormal situations throughout the operational lifecycle, are required. Nevertheless, this is difficult to acquire in a non-destructive approach. In this context, this study investigates an approach to enable a transition from preventive maintenance activities, that are scheduled at predetermined time intervals, into predictive ones. In order to enable such approaches in a cyber-physical production system, a deep learning algorithm is used, allowing for maintenance activities to be planned according to the actual operational status of the machine and not in advance. An autoencoder-based methodology is employed for classifying real-world machine and sensor data, into a set of condition-related labels. Real-world data collected from manufacturing operations are used for training and testing a prototype implementation of Long Short-Term Memory autoencoders for estimating the remaining useful life of the monitored equipment. Finally, the proposed approach is evaluated in a use case related to a steel industry production process.

A Deep Learning Model for Predictive Maintenance in Cyber-Physical Production Systems Using LSTM Autoencoders.

The recent emergence of ubiquitous smart communication devices accelerate people to post the current trending topics in real time as micro blogs, tweets, posts and multimedia content on social media sites along with geographical location tags (geo-tags). Specifically, during recent floods in Tamilnadu 2015, the early warnings about flooded areas emerged to get posted in popular social media with geo-parsed hash tags continuously. In the humanitarian view, the real-time crisis sparked great interest in designing an innovative methodology using big social media data analysis along with supervised machine learning techniques to actuate immediate disaster response and rescue efforts in near future. The proposed system performs disaster tweet collection based on trending disaster hash tags. Our system performs Naive-Bayesian (multinomial) and SSVM classification on collected tweets to identify the severity of the disaster. Based on location-to-interpolation cluster proximity, disaster geographic map is generated for the affected area. Our approach detects the tweets fitted into correct classifier label, and generate an output with detection rate of 79% to 91% of the time. The predicted disaster mapping results are highly accurate up to 89% for real time geo-parsed tweets that matched with actual location at-risk during the flood.

#ChennaiFloods: Leveraging Human and Machine Learning for Crisis Mapping during Disasters Using Social Media

Embedded Bi-directional GRU and LSTMLearning Models to Predict Disasterson Twitter Data

Information entropy based event detection during disaster in cyber-social networks

The advent of sharing sensitive information via Online Social Networks (OSN) has jeopardized the user to the extent that the privacy of millions of OSN users could well be compromised, with their data openly available in the public domain. Evidently, users lack in data privacy and the access control mechanisms available to avoid the risk of disclosure. Therefore a framework that automatically preserves the user privacy to detect sensitive topic and minimize the risk of sensitive information disclosure risk beyond the current privacy sceneries offered by OSN service providers is required. In this paper, we present a three-fold sanitization framework which precisely detects sensitive topics semantically using statistical topic model scheme which incorporates standard knowledge bases for tagging the sensitive topics discovered. The interaction documents from location-of-interest are subjected to SSAR–LDA using Gibbs Sampling to identify sensitive topic clusters with high location entropy. The experimental result shows, (i) the sensitive topic clusters are identified with very high accuracy, (ii) despite the redaction approach, which eliminate the sensitive term, our proposed scheme enhance the privacy preserving policy by replacing the sensitive terms with suitable hierarchical generalization fetched from knowledge bases (iii) the probability of Kullback–Leibler (KL) divergence between sensitive and generalized sanitization terms on Twitter, with negligible information disclosure risk is acceptable, and (iv) the sanitization carried out for 10 sensitive topics, from 4500 user posts of 790 Twitter users, demonstrated high precision and recall, which can be correlated with advanced privacy settings for OSN users in the near future.

Semantics-based sensitive topic diffusion detection framework towards privacy aware online social networks

Social Internet of things (SIoT) has obtained significant attention to address the computational intelligence for handling emergencies which can be sensed through the Internet of smart social things In recent years, humans act as a social sensor to disseminate the information via microblogs in Online Social Networks (OSNs) such as Twitter, Weibo At times, the reaction to certain microblogging prompts thousands of people to rethink and impulsively react on that which can handle humanity unrest situations during earthquakes, floods, Tsunamis, etc In this paper, the streaming microblog tweets are investigated to detect the disaster events for a specified time and location Firstly, the real-time Twitter streaming data is sensed via Apache Flume service agent Secondly, the tweets are preprocessed and filtered tweets are clustered using the microblog-DBSCAN algorithm The tweets belonging to various sliding window from diversified geographic event locations during disasters are aggregated In addition, the crowdsourced photographs are added with geo-tags (location) during the period of analysis stood up as supplementary evidence to detect an event which acts as crisis recovery The dynamic hive queries are executed to filter location-level tweets for the analysis In contrast to conventional approaches which mainly focus the microblog textual content, we incorporate significant metadata features, namely photographs and its geo-tag, to precisely identify the events in real time

A. Bhuvaneswari

Papers

#ChennaiFloods: Leveraging Human and Machine Learning for Crisis Mapping during Disasters Using Social Media

Embedded Bi-directional GRU and LSTMLearning Models to Predict Disasterson Twitter Data

Information entropy based event detection during disaster in cyber-social networks

Semantics-based sensitive topic diffusion detection framework towards privacy aware online social networks

Social IoT-Enabled Emergency Event Detection Framework Using Geo-Tagged Microblogs and Crowdsourced Photographs