Meta learning algorithms for few-shot video recognition use complex, episodic training but they often fail to learn effective feature representations. In contrast, we propose a new and simpler few-shot video recognition method that does not use meta-learning, but its performance compares well with the best meta-learning proposals. Our new few-shot video classification pipeline consists of two distinct phases. In the pre-training phase, we learn a good video feature extraction network that generates a feature vector for each video. After a sparse sampling strategy selects frames from the video, we generate a video feature vector from the sampled frames. Our proposed video feature extractor network, which consists of an image feature extraction network followed by a new transformer encoder, is trained end-to-end by including a classifier head that uses cosine similarity layer instead of the traditional linear layer to classify a corpus of labeled video examples. Unlike prior work in meta learning, we do not use episodic training to learn the image feature vector. Also, unlike prior work that averages frame-level feature vectors into a single video feature vector, we combine individual frame-level feature vectors by using a new Transformer encoder that explicitly captures the key, temporal properties in the sequence of sampled frames. End-to-end training of the video feature extractor ensures that the proposed Transformer encoder captures important temporal properties in the video, while the cosine similarity layer explicitly reduces the intra-class variance of videos that belong to the same class. Next, in the few-shot adaptation phase, we use the learned video feature extractor to train a new video classifier by using the few available examples from novel classes. Results on SSV2-100 and Kinetics-100 benchmarks show that our proposed few-shot video classifier outperforms the meta-learning-based methods and achieves the best state-of-the-art accuracy. We also show that our method can easily discern between actions and their inverse (for example, picking something up vs. putting something down), while prior art, which averages image feature vectors, is unable to do so.

With the growth of 5G, Internet of Things (IoT), edge computing and cloud computing technologies, the infrastructure (compute and network) available to emerging applications (AR/VR, autonomous driving, industry 4.0, etc.) has become quite complex. There are multiple tiers of computing (IoT devices, near edge, far edge, cloud, etc.) that are connected with different types of networking technologies (LAN, LTE, 5G, MAN, WAN, etc.). Deployment and management of applications in such an environment is quite challenging. In this paper, we propose ROMA, which performs resource orchestration for microservices-based 5G applications in a dynamic, heterogeneous, multi-tiered compute and network fabric. We assume that only application-level requirements are known, and the detailed requirements of the individual microservices in the application are not specified. As part of our solution, ROMA identifies and leverages the coupling relationship between compute and network usage for various microservices and solves an optimization problem in order to appropriately identify how each microservice should be deployed in the complex, multi-tiered compute and network fabric, so that the end-to-end application requirements are optimally met. We implemented two real-world 5G applications in video surveillance and intelligent transportation system (ITS) domains. Through extensive experiments, we show that ROMA is able to save up to 90%, 55% and 44% compute and up to 80%, 95% and 75% network bandwidth for the surveillance (watchlist) and transportation application (person and car detection), respectively. This improvement is achieved while honoring the application performance requirements, and it is over an alternative scheme that employs a static and overprovisioned resource allocation strategy by ignoring the resource coupling relationships.

In Video Analytics Pipelines (VAP), Analytics Units (AUs) such as object detection and face recognition running on remote servers critically rely on surveillance cameras to capture high-quality video streams in order to achieve high accuracy. Modern IP cameras come with a large number of camera parameters that directly affect the quality of the video stream capture. While a few of such parameters, e.g., exposure, focus, white balance are automatically adjusted by the camera internally, the remaining ones are not. We denote such camera parameters as non-automated (NAUTO) parameters. In this paper, we first show that environmental condition changes can have significant adverse effect on the accuracy of insights from the AUs, but such adverse impact can potentially be mitigated by dynamically adjusting NAUTO camera parameters in response to changes in environmental conditions. We then present CamTuner, to our knowledge, the first framework that dynamically adapts NAUTO camera parameters to optimize the accuracy of AUs in a VAP in response to adverse changes in environmental conditions. CamTuner is based on SARSA reinforcement learning and it incorporates two novel components: a light-weight analytics quality estimator and a virtual camera that drastically speed up offline RL training. Our controlled experiments and real-world VAP deployment show that compared to a VAP using the default camera setting, CamTuner enhances VAP accuracy by detecting 15.9% additional persons and 2.6%--4.2% additional cars (without any false positives) in a large enterprise parking lot and 9.7% additional cars in a 5G smart traffic intersection scenario, which enables a new usecase of accurate and reliable automatic vehicle collision prediction (AVCP). CamTuner opens doors for new ways to significantly enhance video analytics accuracy beyond incremental improvements from refining deep-learning models.

Identification of people with elevated body temperature can reduce or dramatically slow down the spread of infectious diseases like COVID-19. We present a novel fever-screening system, F3S, that uses edge machine learning techniques to accurately measure core body temperatures of multiple individuals in a free-flow setting. F3S performs real-time sensor fusion of visual camera with thermal camera data streams to detect elevated body temperature, and it has several unique features: (a) visual and thermal streams represent very different modalities, and we dynamically associate semantically-equivalent regions across visual and thermal frames by using a new, dynamic alignment technique that analyzes content and context in real-time, (b) we track people through occlusions, identify the eye (inner canthus), forehead, face and head regions where possible, and provide an accurate temperature reading by using a prioritized refinement algorithm, and (c) we robustly detect elevated body temperature even in the presence of personal protective equipment like masks, or sunglasses or hats, all of which can be affected by hot weather and lead to spurious temperature readings. F3S has been deployed at over a dozen large commercial establishments, providing contact-less, free-flow, real-time fever screening for thousands of employees and customers in indoors and outdoor settings.

The paper introduces Strudel, a development and execution framework for transactional workloads both on SQL and NoSQL systems. Whereas a rich set of benchmarks and performance analysis platforms have been developed for SQLbased systems (RDBMSs), it is challenging for application developers to evaluate both SQL and NoSQL systems for their specific needs. The Strudel framework, which we have released as open-source software, helps such developers (as well as providers of NoSQL stores) to build, customize, and share benchmarks that can run on various SQL/NoSQL systems. We describe Strudel’s architecture and APIs, its components for supporting various NoSQL stores (e.g., HBase, MongoDB), example benchmarks included in the release, and performance experiments to demonstrate usefulness of the framework.

Supporting an online transaction processing (OLTP) workload in a scalable and elastic fashion is a challenging task. Recently, a new breed of scalable systems have shown significant throughput gains by limiting consistency to small units of data called “entity-groups” (e.g., a user's account information stored together with all her emails in an online email service.) Transactions that access the data from only one entity-group are guaranteed of full ACID, but those that access multiple entity-groups are not. Defining entity-groups has direct impact on workload consistency and performance, and doing so for data with a complex schema is very challenging. It is prone to go to extremes - groups that are too fine-grained cause excessive number of expensive distributed transactions while those that are too coarse lead to excessive serialization and performance degradation. It is also difficult to balance conflicting requirements from different transactions. In commercially available entity-group systems, creating entity-groups is usually a manual process, which severely limits the usability of those systems. This paper is the first systematic effort on automating the entity-group design process. Our goal is to build a user-friendly design tool for automatically creating entity-groups based on a given workload and to help users trade consistency for performance in a principled manner. For advanced users, we allow them to provide feedback to the entity-group design and iteratively improve the final output. We demonstrate the effectiveness of our approach with widely used benchmarks. We also present the user experience of a prototype we built.

The demo features Partiqle, a SQL engine over key-value stores as a relational alternative for the recent procedural approaches to support OLTP workloads elastically. Based on our microsharding framework [12], it employs a declarative specification, called transaction classes, of constraints applied on the transactions in a workload. We demonstrate use of a transaction class in design and analysis of OLTP workloads. We then demonstrate live-scaling of our fully functioning system on a server cluster.

We present a data management platform in the cloud, CloudDB. The guiding principle of CloudDB’s design is establishing data independence for the applications that need to use diverse underlying data stores that are optimized for varying workload needs and characteristics. The applications should not have to be aware of the physical organization of the data and how the data is accessed. Ideally, an application only needs a logical specification of the data access layer and the data access requests are handled in a declarative way. CloudDB hosts variety of specialized databases that deliver high performance, scalability, and cost efficiency for varying application needs. CloudDB’s API layer is designed in such a way to give data independence to the higher level applications. The goal is to let the clients use just a simple, standard, and uniform language API to access data management functions as a service.

The UQBE is a mashup tool for non-programmers that supports query-by-example (QBE) over a schema made up by the user without knowing the schema of the original sources. Based on automated schema matching with uncertainty, the UQBE system returns the best confident results. The system lets the user refine them interactively. A tuple in the query result is associated with lineage that is a boolean formula over schema matching decisions representing underlying conditions on which the corresponding tuple is included in the result. Given binary feedbacks on tuples by the user, which are possibly imprecise, the system solves it as an optimization problem to refine confidence values of matching decisions. The demo features graphical user interaction on the UQBE system, including querying and refinement.

Mashup Feeds is a system that supports integrated web service feeds as continuous queries. We introduce collection-based stream processing semantics to enable information extraction by monitoring source evolution over time.

Caching data by maintaining materialized views typically requires updating the cache appropriately to reflect dynamic source updates. Extensive research has addressed the problem of incremental view maintenance for relational data but only few works have addressed it for semi-structured data. In this paper we address the problem of incremental maintenance of views defined over XML documents using path-expressions. The approach described in this paper has the following main features that distinguish it from the previous works: (1) The view specification language is powerful and standardized enough to be used in realistic applications. (2) The size of the auxiliary data maintained with the views depends on the expression size and the answer size regardless of the source data size.(3) No source schema is assumed to exist; the source data can be any general well-formed XML document. Experimental evaluation is conducted to assess the performance benefits of the proposed approach.

For most Web-based applications, contents are created dynamically based on the current state of a business, such as product prices and inventory, stored in database systems. These applications demand personalized content and track user behavior while maintaining application integrity. Many of such practices are not compatible with Web acceleration solutions. Consequently, although many web acceleration solutions have shown promising performance improvement and scalability, architecting and engineering distributed enterprise Web applications to utilize available content delivery networks remains a challenge. In this paper, we examine the challenge to accelerate J2EE-based enterprise web applications. We list obstacles and recommend some practices to transform typical database-driven J2EE applications to cache friendly Web applications where Web acceleration solutions can be applied. Furthermore, such transformation should be done without modification to the underlying application business logic and without sacrificing functions that are essential to e-commerce. We take the J2EE reference software, the Java PetStore, as a case study. By using the proposed guideline, we are able to cache more than 90% of the content in the PetStore and scale up the Web site more than 20 times.

With the wide availability of content delivery networks, many e-commerce Web applications utilize edge cache servers to cache and deliver dynamic contents at locations much closer to users, avoiding network latency. By caching a large number of dynamic content pages in the edge cache servers, response time can be reduced, benefiting from higher cache hit rates. However this is achieved at the expense of higher invalidation cost. On the other hand, a higher invalidation cost leads to a longer invalidation cycle (time to perform the invalidation check on the pages in caches) at the expense of freshness of cached dynamic content. In this paper we propose a freshness-driven adaptive dynamic content caching technique, which monitors response time and invalidation cycle length and dynamically adjusts caching policies. We have implemented the proposed technique within NECs CachePortal Web acceleration solution. The experimental results show that the proposed technique consistently maintains the best content freshness to users. The experimental results also show that even a Web site with dynamic content caching enabled can further benefit from deployment of our solution with improvement of its content freshness up to 10 times especially during heavy traffic.

Wide-area database replication technologies and the availability of content delivery networks allow Web applications to be hosted and served from powerful data centers. This form of application support requires a complete Web application suite to be distributed along with the database replicas. A major advantage of this approach is that dynamic content is served from locations closer to users, leading into reduced network latency and fast response times. However, this is achieved at the expense of overheads due to (a) invalidation of cached dynamic content in the edge caches and (b) synchronization of database replicas in the data center. These have adverse effects on the freshness of delivered content. In this paper, we propose a freshness-driven adaptive dynamic content caching, which monitors the system status and adjusts caching policies to provide content freshness guarantees. The proposed technique has been intensively evaluated to validate its effectiveness. The experimental results show that the freshness-driven adaptive dynamic content caching technique consistently provides good content freshness. Furthermore, even a Web site that enables dynamic content caching can further benefit from our solution, which improves content freshness up to 7 times, especially under heavy user request traffic and long network latency conditions. Our approach also provides better scalability and significantly reduced response times up to 70% in the experiments.