How much can pruning algorithms teach us about the fundamentals of learning representations in neural networks? A lot, it turns out. Neural network model compression has become a topic of great interest in recent years, and many different techniques have been proposed to address this problem. In general, this is motivated by the idea that smaller models typically lead to better generalization. At the same time, the decision of what to prune and when to prune necessarily forces us to confront our assumptions about how neural networks actually learn to represent patterns in data. In this work we set out to test several long-held hypotheses about neural network learning representations and numerical approaches to pruning. To accomplish this we first reviewed the historical literature and derived a novel algorithm to prune whole neurons (as opposed to the traditional method of pruning weights) from optimally trained networks using a second-order Taylor method. We then set about testing the performance of our algorithm and analyzing the quality of the decisions it made. As a baseline for comparison we used a first-order Taylor method based on the Skeletonization algorithm and an exhaustive brute-force serial pruning algorithm. Our proposed algorithm worked well compared to a first-order method, but not nearly as well as the brute-force method. Our error analysis led us to question the validity of many widely-held assumptions behind pruning algorithms in general and the trade-offs we often make in the interest of reducing computational complexity. We discovered that there is a straightforward way, however expensive, to serially prune 40-70\% of the neurons in a trained network with minimal effect on the learning representation and without any re-training.

Existing speech technology tends to be poorly suited for young children at play, both because of their age-specific pronunciation and because they tend to play together, making overlapping speech and side discussions about the play itself ubiquitous. We report the performance of an autonomous, multi-keyword spotter that has been trained and tested on data from a multi-player game designed to focus on these issues. In Mole Madness, children laugh, yell, speak at the same time, make side comments and even invent their own forms of keywords to control a virtual on-screen character. Within this challenging language environment, the system achieves 94 % overall recall and 85 % overall accuracy, providing child-child and child-robot pairs with responsive play in a rapid-paced game. This technology can enable others to create novel multi-party interactions for entertainment where a limited number of keywords has to be recognized.

How much can pruning algorithms teach us about the fundamentals of learning representations in neural networks? A lot, it turns out. Neural network model compression has become a topic of great interest in recent years, and many different techniques have been proposed to address this problem. In general, this is motivated by the idea that smaller models typically lead to better generalization. At the same time, the decision of what to prune and when to prune necessarily forces us to confront our assumptions about how neural networks actually learn to represent patterns in data. In this work we set out to test several long-held hypotheses about neural network learning representations and numerical approaches to pruning. To accomplish this we first reviewed the historical literature and derived a novel algorithm to prune whole neurons (as opposed to the traditional method of pruning weights) from optimally trained networks using a second-order Taylor method. We then set about testing the performance of our algorithm and analyzing the quality of the decisions it made. As a baseline for comparison we used a first-order Taylor method based on the Skeletonization algorithm and an exhaustive brute-force serial pruning algorithm. Our proposed algorithm worked well compared to a first-order method, but not nearly as well as the brute-force method. Our error analysis led us to question the validity of many widely-held assumptions behind pruning algorithms in general and the trade-offs we often make in the interest of reducing computational complexity. We discovered that there is a straightforward way, however expensive, to serially prune 40-70% of the neurons in a trained network with minimal effect on the learning representation and without any re-training.

We report results for an online multi-keyword spotter in a game that contains overlapping speech, off-task side talk, and keyword forms that vary in completeness and duration. The spotter trained on a data set of 62 children, and expectations for online performance were established by 10-fold cross-validation on that corpus. We compare the post hoc data to the recognizer’s performance online in a study in which 24 new children played with the real-time system. The online system showed a non-significant decline in accuracy which could be traced to trouble understanding the jump keyword and the pre-dominance of younger children in the new cohort. However, children adjusted their behavior to compensate, and the overall performance and responsiveness of the online system resulted in engaging and enjoyable gameplay.

We report on our system used in the TRECVID 2015 Multimedia Event Detection (MED) task. On the MED task, the CMU team submitted runs in the Semantic Query (SQ) and 10Ex settings. The proposed system is essentially the same as our MED 2014 system. 1.1 MED System Description On the MED task, the CMU team uses the MED 2014 [1] system which has enabled the system to achieve good performance in the 000Ex and 010Ex settings. Furthermore, our system is very efficient in that it can complete Event Query Generation (EQG) in 16 minutes and Event Search (ES) over 200,000 videos in less than 5 minutes on a single workstation. Please see [1] for the detailed system for the details about our 000Ex and 010Ex runs. The CMU system utilizes multiple modalities, classifiers and fusion methods to perform Multimedia Event Detection. The multiple modalities include visual, audio and text modalities. For 10Ex, two classifiers were used: linear SVM and linear regression. The fusion method used for 010Ex is the Multistage Hybrid Late Fusion, which is a combination of many different fusion algorithms. For the 000Ex runs, we utilize concept detection results from 3000 concept detectors during the SQG and ES stage. We submitted four runs for this year's PS condition: CMU_MED15_MED15EvalFull_PS_10Ex_MED_p-baseline_1: The baseline 10Ex system similar to our 2014 system (using same set of features). CMU_MED15_MED15EvalFull_PS_0Ex_MED_p-expert_1: The 0Ex system using the manual queries selected by experts (using the same queries in our 2014 system). CMU_MED15_MED15EvalFull_PS_0Ex_MED_c-autosqg_1: The 0Ex system using the automatically generated queries. The automatic query generation process is detailed in [3]. CMU_MED15_MED15EvalFull_PS_0Ex_MED_c-autosqgvisualonly_1: The 0Ex system using the automatically generated queries in [3] with only visual features (not including ASR and OCR). 1.2 Hardware Description We utilize the following hardware for metadata generation: 1. PSC Blacklight cluster 100 nodes, each with 4 Intel(R) Xeon(R) CPU E5620 2.40 GHz CPUs (4 cores), 128 GB RAM. Lustre fistributed filesystem, where we used around 50TB. 2. Rocks cluster 20 nodes, each with 2 Intel XEON E5649 2.53 GHz CPUs (6 cores), 64 GB RAM 4 nodes, each with 4 Intel XEON E5-2660 2.20 GHz CPUs (8 cores), 128 GB RAM. 3 nodes, each with 4 NVIDIA TESLA K20 GPUs. 2 data servers, 30TB each For 10Ex event search, we use: 1 Intel(R) Xeon(R) CPU E5-2640 2.50 GHz CPU (12 cores), 128GB RAM, 4 NVIDIA TESLA K20s (2496 cores each), SSD RAID with 4TB storage. For 0Ex event search, we use: 1 Intel(R) Xeon(R) CPU E5649 @ 2.53GHz, 64GB RAM, with a 256GB non-SSD Hard Disk. 1.3 System Performance We report our performance on MED15EvalFull Pre-Specified Events. Runs (MED15EvalFull) Performance MAP% iP10 iP50 infAP200 000Ex autosqgvisualonly 6.4 0.13 0.135 0.0611 000Ex autosqg 7.8 0.2 0.188 0.1005 000Ex expert 15.1 0.38 0.307 0.2137 010Ex baseline 19.2 0.495 0.394 0.2376 Runs (MED15EvalSub) Performance MAP% iP10 iP50 infAP200 000Ex autosqgvisualonly 9.7 0.19 0.148 0.0895 000Ex autosqg 11.0 0.245 0.174 0.1223 000Ex expert 20.6 0.39 0.285 0.246 010Ex baseline 25.5 0.515 0.343 0.2882 Acknowledgments This work used the Extreme Science and Engineering Discovery Environment (XSEDE), which is supported by National Science Foundation grant number OCI-1053575. Specifically, it used the Blacklight system at the Pittsburgh Supercomputing Center (PSC). This work was supported in part by the US Department of Defense, U. S. Army Research Office (W911NF-13-1-0277) and by the National Science Foundation under Grant No. IIS-1251187. The U.S. Government is authorized to reproduce and distribute reprints for Governmental purposes notwithstanding any copyright annotation thereon. Disclaimer: The views and conclusions contained herein are those of the authors and should not be interpreted as necessarily representing the official policies or endorsements, either expressed or implied, of ARO, the National Science Foundation or the U.S. Government. References [1]$ Yu,$Shoou+I.,$et$al.$"Informedia@$TRECVID$2014$MED$and$MER."$NIST$TRECVID$Video$Retrieval$ Evaluation$Workshop.$2014.$ [2]$ Paul$Over$and$George$Awad$and$Martial$Michel$and$Jonathan$Fiscus$and$Wessel$Kraaij$and$Alan$ F.$Smeaton$and$Georges$Quéenot$and$Roeland$Ordelman.$TRECVID$2015$++$An$Overview$of$the$ Goals,$Tasks,$Data,$Evaluation$Mechanisms$and$Metrics.$Proceedings$of$TRECVID$2015,$2015$ [3]$ Lu$Jiang,$Shoou+I$Yu,$Teruko$Mitamura,$Alexander$Hauptmann.$Bridging$the$Ultimate$Semantic$ Gap:$A$Semantic$Search$Engine$for$Internet$Videos.$In$ACM$International$Conference$on$ Multimedia$Retrieval$(ICMR),$2015.$ $ $ $

Renewable energy systems are an increasingly popular way to generate electricity. As with any new technological paradigm, new challenges have emerged which are unique to the utilization of renewable energy systems. One of these challenges in particular is the development of effective monitoring technologies to compensate for the decentralized nature of remote power generation. This project details the development of an open-source monitoring system for remote solar power systems. The problem space that this project is specifically concerned with deals with the reduction of cost and the use of open platforms to make solar monitoring viable in developing countries where both the resources and general knowledge required to undertake such efforts are particularly scarce. Currently, solar monitoring technologies are expensive, limited in their application, and for the most part proprietary. It is arguable that such systems can be developed using non-customized hardware and open-source software that can be obtained and run anywhere in the world. This project is one such argument. This proof of concept is sufficient to show that solar remote monitoring is neither expensive nor particularly cumbersome to implement and thus warrants further investigation and development by the open source community.

One of the main challenges in combating the spread of the Ebola outbreak in West Africa is a lack of effective public health education among affected populations in Guinea, Sierra Leone, and Liberia.Difficulties include resistance to official sources of information, mistrust of government, cultural norms, linguistic barriers, and illiteracy.In this paper we describe the development and initial deployment of a voice-based, multilingual mobile phone application to spread reliable public health information about Ebola via peer-to-peer sharing.Our hypothesis is that we can overcome mistrust and disseminate important health information via the power of social learning and suggestion from friends, family, and local communities.In collaboration with partners on the ground in Conakry, Guinea, we have launched two parallel mobile phone services known as Polly Game and Polly Health to enable message sharing in several Guinean languages.We discuss a variety of strategies we have tried to encourage the spread of the application and data on uptake to date. Operation in Guinea

We report on our system used in the TRECVID 2014 Multimedia Event Detection (MED) and Multimedia Event Recounting (MER) tasks. On the MED task, the CMU team achieved leading performance in the Semantic Query (SQ), 000Ex, 010Ex and 100Ex settings. Furthermore, SQ and 000Ex runs are significantly better than the submissions from the other teams. We attribute the good performance to 4 main components: 1) large-scale semantic concept detectors trained on video shots for SQ/000Ex systems, 2) better features such as improved trajectories and deep learning features for 010Ex/100Ex systems, 3) a novel Multistage Hybrid Late Fusion method for 010Ex/100Ex systems and 4) improved reranking methods for Pseudo Relevance Feedback for 000Ex/010Ex systems. On the MER task, our system utilizes a subset of features and detection results from the MED system from which the recounting is then generated. Recounting evidence is presented by selecting the most likely concepts detected in the salient shots of a video. Salient shots are detected by searching for shots which have high response when predicted by the video level event detector.

Renewable energy systems are an increasingly popular way to generate electricity around the world. As wind and solar technologies gradually begin to supplant the use of fossil fuels as preferred means of energy production, new challenges are emerging which are unique to the experience of decentralized power generation. One such challenge is the development of effective monitoring technologies to relay diagnostic information from remote energy systems to data analysis centers. The ability to easily obtain, synthesize, and evaluate data pertaining to the behavior of a potentially vast number of individual power sources is of critical importance to the maintainability of the next generation of intelligent grid infrastructure. However, the application space of remote monitoring extends well beyond this. This paper details the development and implementation of an open-source monitoring framework for remote solar energy systems. The necessity for such a framework to be open is much better understood when considered through the lens of the theoretical potential for remote monitoring technologies in developing countries. The United States and other industrialized nations in the so-called 'first world' are likely to be slow to seriously adopt renewable energy on account of the massive investment and infrastructural changes required for its integration into the existing electrical grid. In countries where grid infrastructure is generally inadequate or nonexistent, this barrier is far less of a concern, and renewable energy technologies are viewed more as an enabling tool for progress than as a disruptive and expensive technological tangent. In this context as well, remote monitoring has a role to play.

Informedia@TRECVID2014: Semnatic Indexing Report Title We report on our system used in the TRECVID 2014 Semantic Indexing (SIN) task. We highlight the following new components: 1) self-paced learning pipeline for concept training, 2)dense trajectory with fisher vector encoding, 3) multi-modal pseudo relevance feedback for final results reranking and 4) deep convolutional neural networks directly trained on SIN keyframes. With the help of the above components, we were ranked top 3 among all type A runs (using only TRECVID IACC training data). Conference Name: TRECVID Video Retrieval Evaluation Workshop Conference Date: November 10, 2014 Informedia@TRECVID 2014 M E D a nd M E R Shoou-I Yu, Lu Jiang, Zhongwen Xu, Zhenzhong Lan, Shicheng Xu, Xiaojun Chang, Xuanchong Li, Zexi Mao, Chuang Gan, Yajie Miao, Xingzhong Du, Yang Cai, Lara Martin, Nikolas Wolfe, Anurag Kumar, Huan Li, Ming Lin, Zhigang Ma, Yi Yang, Deyu Meng, Shiguang Shan, Pinar Duygulu Sahin, Susanne Burger, Florian Metze, Rita Singh, Bhiksha Raj, Teruko Mitamura, Richard Stern and Alexander Hauptmann Carnegie Mellon University 5000 Forbes Ave., Pittsburgh,15213 We report on our system used in the TRECVID 2014 Multimedia Event Detection (MED) and Multimedia Event Recounting (MER) tasks. On the MED task, the CMU team achieved leading performance in the Semantic Query (SQ), 000Ex, 010Ex and 100Ex settings. Furthermore, SQ and 000Ex runs are significantly better than the submissions from the other teams. We attribute the good performance to 4 main components: 1) large-scale semantic concept detectors trained on video shots for SQ/000Ex systems, 2) better features such as improved trajectories and deep learning features for 010Ex/100Ex systems, 3) a novel Multistage Hybrid Late Fusion method for 010Ex/100Ex systems and 4) improved reranking methods for Pseudo Relevance Feedback for 000Ex/010Ex systems. On the MER task, our system utilizes a subset of features and detection results from the MED system from which the recounting is then generated. Recounting evidence is presented by selecting the most likely concepts detected in the salient shots of a video. Salient shots are detected by searching for shots which have high response when predicted by the video level event detector.

In this paper, we describe the concept of an interactive tool which can be employed to build a dialog system to facilitate pronunciation training in situations where only a few minutes of speech in the target language are available. We leverage recent advances in low-resource speech processing, and envision a tool which will help organizations working in the developing world to quickly create initial training lessons for new languages. Development will be possible using mobile devices alone, and without the requirement of significant technical skills or language-specific information. If users are ultimately able to acquire at least rudimentary proficiency in a low-resource language or dialect using our automatic system, they should find it much easier to communicate, establish trust, and build rapport with the local population than without such support. In this paper, we present the operational principles of our approach, describe a proof-of-concept implementation that we are currently developing, and summarize ideas for evaluation at the technical, language-learning, and user interface levels. Author

We report on our system used in the TRECVID 2014 Multimedia Event Detection (MED) and Multimedia Event Recounting (MER) tasks. On the MED task, the CMU team achieved leading performance in the Semantic Query (SQ), 000Ex, 010Ex and 100Ex settings. Furthermore, SQ and 000Ex runs are significantly better than the submissions from the other teams. We attribute the good performance to 4 main components: 1) our large-scale semantic concept detectors trained on video shots for SQ/000Ex systems, 2) better features such as improved trajectories and deep learning features for 010Ex/100Ex systems, 3) a novel Multistage Hybrid Late Fusion method for 010Ex/100Ex systems and 4) our developed reranking methods for Pseudo Relevance Feedback for 000Ex/010Ex systems. On the MER task, our system utilizes a subset of features and detection results from the MED system from which the recounting is then generated. Recounting evidence is presented by selecting the most likely concepts detected in the salient shots of a video. Salient shots are detected by searching for shots which have high response when predicted by the video level event detector.

Surveillance Event Detection Semantic event detection in the huge amount of surveillance video in both retrospective and real-time styles is essential to a variety of higher-level applications in the public security. In TRECVID 2010, to overcome the limitations of the traditional human action analysis method with human detection/tracking and domain knowledge, we evaluate the general framework for multiple human behaviors modeling with the philosophy of bag of spatiotemporal feature (BoSTF). The brief