On stabbing queries for generalized longest repeat

In the context of automated analysis of eye fundus images, it is an important common fallacy that prior works achieve very high scores in segmentation of lesions, and that fallacy is fueled by some reviews reporting very high scores, and perhaps some confusion with terms. A simple analysis of the detail of the few prior works that really do segmentation reveals scores between 7% and 70% in sensitivity for 1 FPI. That is clearly sub-par with medical doctors trained to detect signs of Diabetic Retinopathy, since they can distinguish well the contours of lesions in Eye Fundus Images (EFI). Still, a full segmentation of lesions could be an important step for both visualization and further automated analysis using rigorous quantification or areas and numbers of lesions to better diagnose. I discuss what prior work really does, using evidence-based analysis, and confront with segmentation networks, comparing on the terms used by prior work to show that the best performing segmentation network outperforms those prior works. I also compare architectures to understand how the network architecture influences the results. I conclude that, with the correct architecture and tuning, the semantic segmentation network improves up to 20 percentage points over prior work in the real task of segmentation of lesions. I also conclude that the network architecture and optimizations are important factors and that there are still important limitations in current work.

Existing machine learning algorithms for human information inference are typically data-driven models trained on carefully labeled datasets. Given the significant labeling effort, traditional pure data-driven approaches are challenging to implement for emerging smart applications requiring long-term finer-grained information. Taking activities of daily life (ADL) tracking for elders as an example, prior work mostly focused on context-level information learning such as cooking and cleaning. [8]. However, new applications such as evaluating elders' cognitive impairments progress by tracking their ADL engagement requires finer-grained, i.e., action-level information [7]. In practice, labeling the day-length data at such granularity can be very expensive and requires a lot of human efforts [9]. My research focuses on the inference problems in the scope of human physical condition monitoring and activity recognition with limited labeled data. To alleviate the effort of labeling large amounts of data, prior works on semi-supervised learning combine a small amount of labeled data with a large amount of unlabeled data to train the model. However, as the label granularity (number of classes) increasing, the difficulty to distinguish nuance distinctions between finer-grained classes escalates as well. This makes training a robust semi-supervised model for finer-grained classification with less labels difficult if not impossible. Fortunately, coarse-grained (context-level) labels is usually available or cheaper to obtain in practice. In this case, the multi-granularity hierarchy between finer and coarse labels follows the aggregation relation defined in [5]. This hierarchical relation can be leveraged in the tasks of inferring finer-grained information. In addition, it is illustrated by the previous study that co-located multi modality sensing systems capture complementary aspects of the same event [6]. The research question I focus on is how to infer finer-grained human information with coarse-grained labeled data leveraging complementary multi-modal sensing? I target three directions: 1) a cross-granularity semi-supervised setting: how to utilize coarse-grained labeled data with a small amount of finer-grained labeled data to infer finer-grained human information, 2) cross-granularity relationship learning: how to learn the multi-granularity class hierarchy from data and further help the finer-grained human information acquisition, 3) enhancing inference granularity by leveraging multi-model sensing: how to leverage the complimentary co-located multiple sensing modalities to accurately infer finer-grained human information?

Traditional polarization gratings (PGs) have been studied with increasing intensity since 2005, in part because they can manifest 100% single-order diffraction efficiency and strong sensitivity to input polarization, in both theory and practice. They can be made using patterned anisotropic materials (e.g., liquid crystals) or nanostructures (e.g., metasurfaces). Nearly every prior work on traditional PGs has implemented a linear spatial phase-shift that is either continuous or which samples the 2π phase period with multiple (≥ 4) discrete phase levels. As far as we know, only two prior works (Bhandari et al, Phys. Rep. 281 (1997); and Wang et al, Appl. Phys. Lett. 108 (2016)) have considered the circumstance when the phase is sampled with exactly two phase levels, with π radians between them. We call this a Binary PG (Bin-PG). In this work, we apply Jones calculus and the small angle (i.e., paraxial) approximation to derive the fundamental optical behavior of Bin-PGs: far-field efficiencies, input polarization sensitivity, and output polarizations. We show that Bin-PGs manifest properties that are a compelling and unique mixture of both traditional (non-binary) PGs and standard diffraction gratings (e.g., surface-relief-gratings (SRGs)). Like non-binary PGs, their output polarization is often different than the input and diffraction efficiencies are dependent on the effective retardation of the film or surface. However, like SRGs, they show a maximum of 80% total first-order efficiency and are insensitive to input polarization.

Virtual metrology (VM) is an enabling technology capable of performing virtual inspection on the metrology quality of wafers. Instead of physically acquiring the metrology measurements, VM applies conjecture models on the process data of wafers to estimate the measurements of the targeted metrology variables. Prior works on overlay VM system utilized fault detection and classification (FDC) data as the process data for the conjecture models. Hence, when FDC data are unavailable owing to FDC system enhancement works, FDC-based VM models would be rendered inefficacious. During such events, a competent VM system using a different modeling approach is required to sustain the production line until FDC data resumes availability and FDC-based VM reaches production state. Motivated by a real-world production environment of a 200mm semiconductor manufacturing plant (fab), a novel wafer lot-level modeling approach for overlay VM was proposed in our prior work. Using the proposed modeling, a smart sampling scheme was also designed in the same work. The smart sampling scheme consists of two conjecture tasks, with the first task classifies the overlay quality of the wafers, and the second task estimates the overlay errors of the wafers classified with normal overlay quality. The abnormal ones are diverted to the physical metrology station. In this paper, the implementation of a smart sampling system, C2O, using the designed scheme and its experimental results are presented. The experimental results showed that C2O is capable to achieve a true positive rate (TPR) of 71.34% for the classification task and mean absolute scaled error (MASE) of 9.59 for the regression task. The obtained results set the baseline to measure the efficacy of future enhancement works, which have been enlisted and underway to augment the performance of the system so that its competency meets the requirements of real fab.

We study distributed causal shared memory (or distributed read/write objects) in the client-server model over asynchronous message-passing networks in which some servers may suffer Byzantine failures. Since Ahamad et al. proposed causal memory in 1994, there have been abundant research on causal storage. Lately, there is a renewed interest in enforcing causal consistency in large-scale distributed storage systems (e.g., COPS, Eiger, Bolt-on). However, to the best of our knowledge, the fault-tolerance aspect of causal memory is not well studied, especially on the tight resilience bound. In our prior work, we showed that 2 f+1 servers is the tight bound to emulate crash-tolerant causal shared memory when up to f servers may crash. In this paper, we adopt a typical model considered in many prior works on Byzantine-tolerant storage algorithms and quorum systems. In the system, up to f servers may suffer Byzantine failures and any number of clients may crash. We constructively present an emulation algorithm for Byzantine causal memory using 3 f+1 servers. We also prove that 3 f+1 is necessary for tolerating up to f Byzantine servers. In other words, we show that 3 f+1 is a tight bound. For evaluation, we implement our algorithm in Golang and compare their performance with two state-of-the-art fault-tolerant algorithms that ensure atomicity in the Google Cloud Platform.

An optimal control theory for designing a controller, that selects unique control actions is developed for controlling discrete event systems (DESs). Prior work on supervisor synthesis problems studied the design of supervisors for satisfying qualitative specifications within the framework of the DESs. In this paper we define and investigate the synthesis of an optimal controller in the sense that, regardless the state of the plant, the worst cost of all surviving paths from that state to the set of frontier marked states is minimized. Our work differs from prior works on optimal control of DESs in that the controller selects a unique controllable event to be executed per state, rather than a set of admissible controllable events. An efficient algorithm is provided to solve this problem under complete observation.

In order to provide a streaming service of 3-D spatial objects to the mobile clients on a street, we propose a new query type, called continuous perspective query. The perspective query differs from conventional spatial queries in that the levels of details (LOD) of query results depend on the distance between the query point and spatial objects. The objects in the vicinity are to be provided with a higher LOD than those far from the query point. We are dealing with continuous queries, and the LODs of the results are also changing according to the distance from the mobile query point. The LOD of the result object changes from a low LOD to a high LOD as the mobile query point approaches to the object. In this paper, we also propose a processing method for continuous perspective query to reduce the processing cost at a server and communication cost between the server and the mobile clients.

DRAM cells require periodic refreshing to preserve data. In JEDEC DDRx devices, a refresh operation is performed via an auto-refresh command, which refreshes multiple rows in multiple banks simultaneously. The internal implementation of auto-refresh is completely opaque outside the DRAM --- all the memory controller can do is to instruct the DRAM to refresh itself --- the DRAM handles all else, in particular determining which rows in which banks are to be refreshed. This is in conflict with a large body of research on reducing the refresh overhead, in which the memory controller needs fine-grained control over which regions of the memory are refreshed. For example, prior works exploit the fact that a subset of DRAM rows can be refreshed at a slower rate than other rows due to access rate or retention period variations. However, such row-granularity approaches cannot use the standard auto-refresh command, which refreshes an entire batch of rows at once and does not permit skipping of rows. Consequently, prior schemes are forced to use explicit sequences of activate (ACT) and precharge (PRE) operations to mimic row-level refreshing. The drawback is that, compared to using JEDEC's auto-refresh mechanism, using explicit ACT and PRE commands is inefficient, both in terms of performance and power. In this paper, we show that even when skipping a high percentage of refresh operations, existing row-granurality refresh techniques are mostly ineffective due to the inherent efficiency disparity between ACT/PRE and the JEDEC auto-refresh mechanism. We propose a modification to the DRAM that extends its existing control-register access protocol to include the DRAM's internal refresh counter. We also introduce a new "dummy refresh" command that skips refresh operations and simply increments the internal counter. We show that these modifications allow a memory controller to reduce as many refreshes as in prior work, while achieving significant energy and performance advantages by using auto-refresh most of the time.

Collaborative crowd computing, e.g., human computation and crowdsourcing, involves a team of workers jointly solving tasks of varying difficulties. In such settings, the ability to manage the workflow based on workers' skills and task strains can improve output quality. However, many practical systems employ a simple additive scoring scheme to measure worker performance, and do not consider the task difficulty or worker interaction. Some prior works have looked at ways of measuring worker performance or task difficulty in collaborative settings, but usually assume sophisticated models. In our work, we address this question by taking a competitive perspective and leveraging the vast prior work on competitive games. We adapt TrueSkill's standard competitive model by treating the task as a fictitious worker that the team of humans jointly plays against. We explore two fast online approaches to estimate the worker and task ratings: (1) an ELO rating system, and (2) approximate inference with the Expectation Propagation algorithm. To assess the strengths and weaknesses of the various rating methods, we conduct a human study on Amazon's Mechanical Turk with a simulated ESP game. Our experimental design has the novel element of pairing a carefully designed bot with human workers; these encounters can be used, in turn, to generate a larger set of simulated encounters, yielding more data. Our analysis confirms that our ranking scheme performs consistently and robustly, and outperforms the traditional additive scheme in terms of predicted accuracy. © 2019 IW3C2 (International World Wide Web Conference Committee), published under Creative Commons CC-BY 4.0 License.

Source code commenting is a common practice to improve code comprehension in software development. While comments often consist of descriptive natural language, surprisingly, there exists a non-trivial portion of comments that are actually code statements, i.e., commented-out code (CO code), even in well-maintained software systems. Commented-out code practice is rarely studied and often excluded in prior studies on comments due to its irrelevance to natural language. When being openly discussed, CO practice is generally considered a bad practice. However, there is no prior work to assess the nature (prevalence, evolution, motivation, and necessity of utilization) of CO code practice. In this paper, we perform the first study to understand CO code practice. Inspired by prior works in comment analysis, we develop automated solutions to identify CO code and track its evolution in development history. Through analyzing six open-source projects of different sizes and from diverse domains, we find that CO code practice is non-trivial in software development, especially in the early phase of development history, e.g., up to 20% of the commits involve CO code practice. We observe common evolution patterns of CO code and find that developers may uncomment and comment code more frequently than expected, e.g., 10% of the CO code practices have been uncommented at least once. Through a manual analysis, we identify the common reasons that developers adopt CO code practices and reveal maintenance challenges associated with CO code practices.

Deep neural networks (DNNs) are being widely used in various computer vision tasks as they can achieve very high accuracy. However, the large number of parameters employed in DNNs can result in long inference times for vision tasks, thus making it even more challenging to deploy them in the compute- and memory-constrained mobile/edge devices. To boost the inference of DNNs, some existing works employ compression (model pruning or quantization) or enhanced hardware. How-ever, most prior works focus on improving model structure and implementing custom accelerators. As opposed to the prior work, in this paper, we target the video data that are processed by edge devices, and study the similarity between frames. Based on that, we propose two runtime approaches to boost the performance of the inference process, while achieving high accuracy.Specifically, considering the similarities between successive video frames, we propose a frame-level compute reuse algorithm based on the motion vectors of each frame. With frame-level reuse, we are able to skip 53% of frames in inference with negligible overhead and remain within less than 1% mAP (accuracy) drop for the object detection task. Additionally, we implement a partial inference scheme to enable region/tile-level reuse. Our experiments on a representative mobile device (Pixel 3 Phone) show that the proposed partial inference scheme achieves 2 × speedup over the baseline approach that performs full inference on every frame. We integrate these two data reuse algorithms to accelerate the neural network inference and improve its energy efficiency. More specifically, for each frame in the video, we can dynamically select between (i) performing a full inference, (ii) performing a partial inference, or (iii) skipping the inference altogether. Our experimental evaluations using six different videos reveal that the proposed schemes are up to 80% (56% on average) energy efficient and 2.2× performance efficient compared to the conventional scheme, which performs full inference, while losing less than 2% accuracy. Additionally, the experimental analysis indicates that our approach outperforms the state-of-the-art work with respect to accuracy and/or performance/energy savings.

Efficiently answer top-k queries on typed intervals

Data-Driven Approaches to Maximize the Impact of Pediatric Clinical Trials.

Developed from Nearest Neighbor (NN) queries, Aggregate Nearest Neighbor (ANN) queries return the object that minimizes an aggregate distance function with respect to a set of query points. Because of the multiple query points, ANN queries are much more complex than NN queries. For optimizing the query processing and improving the query efficiency, many ANN queries algorithms utilizes pruning strategies, with or without an index structure. Obviously, the pruning effect highly depends on the tightness of the bound estimation. In this paper, we figure out a property in vector space and develop some efficient bound estimations for two most popular types of ANN queries. Based on these bounds, we design the indexed and non-index ANN algorithms, and conduct experimental studies. Our algorithms show good performance, especially for high dimensional queries, for both real dataset and synthetic datasets.

In this paper, for the first time, we present the concept of nearest collection (NC) search. Given a set of spatial data points D and a query point q, a nearest collection search retrieves a certain subset c (|c| = k), called collection from D. We formally define a collection as clustered k objects and the nearest collection search problem. Since the brute-force approach of this problem requires large computational cost, we propose two approaches using database techniques to reduce search space. The first approach is the multiple query method which uses existing method (i.e. k-nearest neighbor query) based on normal R-tree. The second approach is the effective NC query processing based on the branch and bound method using an aggregate R-tree (simply aR-tree). Our experimental results show that the efficiency and effectiveness of our proposed approach.