The Convolution Neural Network (CNN) has demonstrated the unique advantage in audio, image and text learning; recently it has also challenged Recurrent Neural Networks (RNNs) with long short-term memory cells (LSTM) in sequence-to-sequence learning, since the computations involved in CNN are easily parallelizable whereas those involved in RNN are mostly sequential, leading to a performance bottleneck. However, unlike RNN, the native CNN lacks the history sensitivity required for sequence transformation; therefore enhancing the sequential order awareness, or position-sensitivity, becomes the key to make CNN the general deep learning model. In this work we introduce an extended CNN model with strengthen position-sensitivity, called PoseNet. A notable feature of PoseNet is the asymmetric treatment of position information in the encoder and the decoder. Experiments shows that PoseNet allows us to improve the accuracy of CNN based sequence-to-sequence learning significantly, achieving around 33-36 BLEU scores on the WMT 2014 English-to-German translation task, and around 44-46 BLEU scores on the English-to-French translation task.

The 8th International Workshop on Business Intelligence for the Real-Time Enterprise (BIRTE) was held on September 1, 2014 in conjunction with the VLDB 2014 Conference in Hangzhou, China. Like in previous years, the workshop was well attended by an engaged audience from both academia and industry, breaking an attendance record in the history of the BIRTE workshop series with more than 80 participants during the keynote session. In addition to the keynote speech, the workshop included two invited industrial talks, and four presentations of peer-reviewed papers covering a wide range of real-time BI topics with an overarching emphasis on big data analytics. The workshop developed as follows: After the official opening of the workshop, two papers were presented one on BI analytics on graph-structured data and another on contextual analysis in temporal databases. The invited industrial talks session that followed included a talk from Microsoft Research about their data processing solutions for complex big data analytics and a talk from HP Labs about a new fault tolerance technique that they developed for distributed stream processing. The afternoon sessions opened with the highlight of the workshop: Dr. C. Mohan’s keynote speech providing a critical survey of the current big data landscape. The workshop closed with a session that consisted of the two remaining papers one on exploratory OLAP over linked data and another on data stream partitioning.

Combining data warehousing and stream processing technologies has great potential in offering low-latency data-intensive analytics. Unfortunately, such convergence has not been properly addressed so far. The current generation of stream processing systems is in general built separately from the data warehouse and query engine, which can cause significant overhead in data access and data movement, and is unable to take advantage of the functionalities already offered by the existing data warehouse systems.In this work we tackle some hard problems in integrating stream analytics capability into the existing query engine. We define an extended SQL query model that unifies queries over both static relations and dynamic streaming data, and develop techniques to extend query engines to support the unified model. We propose the cut-and-rewind query execution model to allow a query with full SQL expressive power to be applied to stream data by converting the latter into a sequence of "chunks", and executing the query over each chunk sequentially, but without shutting the query instance down between chunks for continuously maintaining the application context across the execution cycles as required by sliding-window operators. We also propose the cycle-based transaction model to support Continuous Querying with Continuous Persisting (CQCP) with cycle-based isolation and visibility.We have prototyped our approach by extending the PostgreSQL. This work has resulted in a new kind of tightly integrated, highly efficient system with the advanced stream processing capability as well as the full DBMS functionality. We demonstrate the system with the popular Linear Road benchmark, and report the performance. By leveraging the matured code base of a query engine to the maximal extent, we can significantly reduce the engineering investment needed for developing the streaming technology. Providing this capability on proprietary parallel analytics engine is work in progress.

Support transaction property and fault-tolerance is the key to applying stream processing to industry-scale applications; however the corresponding latency overhead must be minimized for accommodating real-time analytics. This issue has been studied in various contexts. In this work we develop the backtrack failure recovery mechanism to allow a task to roll forward without waiting for acknowledgement from its downstream target tasks in the failure-free case, but to request its upstream source tasks to resend the missing tuples only during failure recovery which is the rare case thus has limited impact on the overall performance. For further reduced latency we extend our solution in another dimension by applying the notion of optimistic checkpointing to stream processing, and propose the Continued stream processing with Window - based Checkpoint and Recovery (CWCR) approach, allowing a task to emit results tuple by tuple continuously but checkpoint in batch, and acknowledge, only once per window (e.g. time window). We also tackle the hard problems found in implementing a transactional layer on-top of an existing stream processing platform. We have implemented the proposed mechanisms on Fontainebleau, the distributed stream analytics infrastructure we built on top of the open-sourced Storm platform. Our experiment results reveal the novelty of the proposed technologies and the feasibility to support fault-tolerance with minimal latency overhead for real-time stream processing.