Full Raw Data Research Articles

Continual Learning From a Stream of APIs.

Continual learning (CL) aims to learn new tasks without forgetting previous tasks. However, existing CL methods require a large amount of raw data, which is often unavailable due to copyright considerations and privacy risks. Instead, stakeholders usually release pre-trained machine learning models as a service (MLaaS), which users can access via APIs. This paper considers two practical-yet-novel CL settings: data-efficient CL (DECL-APIs) and data-free CL (DFCL-APIs), which achieve CL from a stream of APIs with partial or no raw data. Performing CL under these two new settings faces several challenges: unavailable full raw data, unknown model parameters, heterogeneous models of arbitrary architecture and scale, and catastrophic forgetting of previous APIs. To overcome these issues, we propose a novel data-free cooperative continual distillation learning framework that distills knowledge from a stream of APIs into a CL model by generating pseudo data, just by querying APIs. Specifically, our framework includes two cooperative generators and one CL model, forming their training as an adversarial game. We first use the CL model and the current API as fixed discriminators to train generators via a derivative-free method. Generators adversarially generate hard and diverse synthetic data to maximize the response gap between the CL model and the API. Next, we train the CL model by minimizing the gap between the responses of the CL model and the black-box API on synthetic data, to transfer the API's knowledge to the CL model. Furthermore, we propose a new regularization term based on network similarity to prevent catastrophic forgetting of previous APIs. Our method performs comparably to classic CL with full raw data on the MNIST and SVHN datasets in the DFCL-APIs setting. In the DECL-APIs setting, our method achieves 0.97×, 0.75× and 0.69× performance of classic CL on the more challenging CIFAR10, CIFAR100, and MiniImageNet, respectively.

Direct Target Localization With Quantized Measurements in Noncoherent Distributed MIMO Radar Systems

In this paper, a direct target localization algorithm with quantized measurements for non-coherent MIMO systems is proposed. In this system, each receiver transmits a low-bit quantized version of the echo rather than the full raw data to the fusion center. We construct a joint likelihood function based on the low-bit data of each receiver, from which the unknown target position can be directly determined. The Cramer-Rao lower bound (CRLB) is derived to analyze the localization performance of our proposed algorithm. To maximize the localization performance, a CRLB-based objective function is designed to obtain the optimum quantization thresholds. The formulated problem is a high-dimensional and non-convex optimization problem that when solved, determines two types of coupled parameters for the quantization thresholds and the complex-valued scaling coefficients of the signal. We propose a batch gradient descent embedded particle swarm optimization algorithm to solve this problem effectively. Numerical results show that the proposed algorithm delivers superior performance in terms of maximizing the overall localization performance, and the 3-bit quantized algorithm is able to provide performance that is very close to the unquantized algorithm. Experimental data recorded by three small radars are also provided to demonstrate the effectiveness of the proposed algorithm.

The Belle II Raw Data Management System

The Belle II experiment, a major upgrade of the previouse+e−asymmetric collider experiment Belle, is expected to produce tens of petabytes of data per year due to the luminosity increase from the upgraded SuperKEKB accelerator. The distributed computing system of the Belle II experiment plays a key role, storing and distributing data in a reliable way to be easily accessed and analyzed by more than 1000 collaborators. In particular, the Belle II Raw Data Management system has been developed with an aim to upload output files onto grid storage, register them into the file and metadata catalogs, and make two replicas of the full raw data set using the Belle II Distributed Data Management system. It has been implemented as an extension of DIRAC (Distributed Infrastructure with Remote Agent Control) and consists of a database, services, client and monitoring tools, and several agents that treat the data automatically. The first year of data taken with the Belle II full detector has been managed by the Belle II Raw Data Management system successfully. The design, current status, and performance are presented. Prospects for improvements towards the full luminosity data taking are also reviewed.

Extensive selective reporting of quality of life in clinical study reports and publications of placebo-controlled trials of antidepressants.

Selective reporting of trial results is common. To study selective reporting in clinical study reports, company trial registers and publications of quality of life in placebo-controlled trials of antidepressants. We compared clinical study reports of four antidepressants (fluoxetine, duloxetine, paroxetine and sertraline) obtained from two European drug regulators, data from online company registers, and publications received or retrieved from Eli Lilly and GlaxoSmithKline. Pfizer was also contacted but did not provide any publications. We included 15 trials (19,015 pages) and 4717 patients. Six trials had used SF-36, seven EQ-5D and two both instruments. Nine of the 15 CSRs (60%) displayed selective reporting. In the companies' online registers, there was selective reporting for all 15 trials (100%). We received 20 publications from Eli Lilly and retrieved six from the GlaxoSmithKline register. There was selective reporting in 24 of the 26 publications (92%). Despite extensive selective reporting, we found only small differences between placebo and active drugs. Access to the full raw data from clinical trials and to case report forms for all patients are needed to evaluate the effect of antidepressants on quality of life. Regulatory agencies should refuse to approve drugs or new indications based on incomplete reporting.

$L_{1}$ -Regularization-Based SAR Imaging and CFAR Detection via Complex Approximated Message Passing

Synthetic aperture radar (SAR) is a widely used active high-resolution microwave imaging technique that has all-time and all-weather reconnaissance ability. Compared with traditionally matched filtering (MF)-based methods, ${L_{q}}({0 \le q \le 1})$ regularization technique can efficiently improve SAR imaging performance e.g., suppressing sidelobes and clutter. However, conventional $L_{q}$ -regularization-based SAR imaging approach requires transferring the 2-D echo data into a vector and reconstructing the scene via 2-D matrix operations. This leads to significantly more computational complexity compared with MF, and makes it very difficult to apply in high-resolution and wide-swath imaging. Typical $L_{q}$ regularization recovery algorithms, e.g., iterative thresholding algorithm, can improve imaging performance of bright targets, but not preserve the image background distribution well. Thus, image background statistical-property-based applications, such as constant false alarm rate (CFAR) detection, cannot be applied to regularization recovered SAR images. On the other hand, complex approximated message passing (CAMP), an iterative recovery algorithm for $L_{1}$ regularization reconstruction, can achieve not only the sparse estimation of the original signal as typical regularization recovery algorithms but also a nonsparse solution simultaneously. In this paper, two novel CAMP-based SAR imaging algorithms are proposed for raw data and complex radar image data, respectively, along with CFAR detection via the CAMP recovered nonsparse result. The proposed method for raw data can not only improve SAR image performance as conventional $L_{1}$ regularization technique but also reduce the computational cost efficiently. While only when we have MF recovered SAR complex image rather than raw data, the proposed method for complex image data can achieve a similar reconstructed image quality as the regularization-based SAR imaging approach using the full raw data. The most important contribution of this paper is that the proposed CAMP-based methods make CFAR detection based on the regularization reconstruction SAR image possible using their nonsparse scene estimations, which has a similar background statistical distribution as the MF recovered images. The experimental results validated the effectiveness of the proposed methods and the feasibility of the recovered nonsparse images being used for CFAR detection.

Multi-Domain Processing for Micropower Sensors

Multi-Domain Processing for Micropower Sensors

Mislabelled table entries in ADVISE Registry by Petraco and colleagues

Petraco and colleagues1 have presented the “adjusted” accuracy comparing their new index (instantaneous wave-free ratio, iFR) against the gold standard of fractional flow reserve (FFR). However, Table 2 of their manuscript misrepresents theoretical calculations as clinical observations. Misleading column labels hide the fact from the casual reader that many of its numbers are assumed from a model instead of being measured directly. Specifically, in three of the four studies (ADVISE registry, ADVISE study, FFR-PET study) FFR values were measured only once, yet the table makes no distinction among the agreement data in its “Repeated FFR measurements” column. Even for the DEFER study, which actually repeated FFR measurements, the authors did not have access to the full raw data. Similarly, two of the four studies (DEFER, FFR-PET study) never measured iFR, yet the table presents an “observed” agreement between iFR and FFR for all rows. Should not measured values – true observations – be distinguished from assumptions? For the last row of the table (FFR-PET study), this confluence of hypothetical values reaches too far, calculating an “adjusted” iFR accuracy by dividing the iFR-FFR agreement (for a study that never measured iFR at all) by the repeated FFR agreement (for a study that only measured FFR once). Only after careful reading of the methods section can the reader uncover that five of the eight values (>50%) in the “Overall classification agreement” columns of Table 2 are an estimation instead of a measurement. Indeed, each and every “adjusted” iFR accuracy in Table 2 contains at least one component that has been assumed from a model. Therefore, their statement that the so-called adjusted “iFR accuracy is almost identical, ranging from 94% to 96%” follows trivially from the underlying assumptions. To our knowledge, in the peer-reviewed literature only the VERIFY study2 has actually measured both iFR and FFR twice in the same patients. The VERIFY study found superior reproducibility for repeated FFR measurements compared to repeated iFR measurements. Even their proposal to “adjust” the agreement suffers from three statistical shortcomings as we will detail in a future manuscript. First, mathematically it does not estimate the true, underlying agreement between the two variables. Second, it only accounts for variability in FFR while neglecting the variability in iFR measurements. Third, it does not generalise beyond a single repetition, whereas investigators may perform two or even more repeated measurements. Fundamentally, Table 2 by Petraco and colleagues falls short of presenting its contents accurately.

Extending the IceCube DAQ System by Integration of the Generic High-Speed Sorter Module TESS

In the extreme environment of Antarctica at the South Pole, the IceCube experiment, the world's first kilometer-scale neutrino telescope, collects cosmic ray events. IceCube consists of over 5000 digital optical sensor modules (DOMs) deployed on 86 instrumentation lines each extending 2.5 km deep in the antarctic ice. The array of optical modules monitors the Cherenkov light emitted by passing radiation, which, when digitized and timestamped to nanosecond precision, is used as input to sophisticated reconstruction algorithms that determine the direction, energy, and type of the incident cosmic ray event. In order to achieve this goal, the IceCube data acquisition system merges the digital data streams from each photodetector into a single time-ordered list which is presented to online triggers that determine, in realtime, whether or not a given pattern of hits is noise or signal. At the present time, the data provided to the triggers is limited by the performance of sorting and merging algorithms: the 500 Hz raw data rate from each sensor (2.5 MHz array aggregate rate) is beyond the capability of the central sort and merge. The current solution adopted by the IceCube detector is to impose a hardware-based pretrigger coincidence on hits emanating from the DOMs which reduces the rate by a factor of 20. While this pretrigger coincidence has negligible impact on the detector sensitivity for the principal goal of high-energy neutrinos from galactic or extragalactic sources, other low-energy physics searches are affected. This presentation details work done to develop and implement a system, TESS, which is capable of merging the full raw data stream being produced by the IceCube DOMs. TESS is designed as a pipelined architecture with three major modules: server, selector and the client glued together by circular buffers. The three modules runs in only three threads and since the architecture is self synchronizing and uses no data copying maximum performance can be achieved for global sorting of payloads. The TESS sorting architecture was originally designed to provide a globally sorted data stream for triggers targeting low-energy events from annihilation of hypothesized dark-matter particles, however its utility is generalizable to any IceCube trigger which requires inspection of the full data stream. The IceCube online supernova detection system is a notable example. Moreover, the algorithm is generic to any system involving multiple, independently sorted data streams which must be merged into a single sorted data stream.

Investigating post-processing of phased array data for detection and sizing capabilities using incoherent compounding

Advanced non-destructive techniques integrate ultrasonic arrays with array controllers that perform full waveform capture (FWC). FWC is the term used to describe the acquisition of RF waveforms corresponding to specific transmit-receive beams from an array. With FWC, the full waveform is available for post-processing. However, FWC only allows rectification and gating during data acquisition and post-processing. If the correct information is not acquired the first time, inspection cost and time may be compromised. FRD (full raw data) collection works similarly to FWC, acquiring the time-domain responses from the excited phased array transmit-receive combinations. However, the transmit and receive apertures for FRD are much smaller than for FWC - typically one element each - and there are many receive locations for each transmit location. The main advantages of FRD collection are the facility for real-time imaging and to optimise features such as beam steering, aperture shading and focusing during the post-processing stage(1). Compounding, a technique previously used for biomedical applications, has the potential to be applied as a novel technique in NDT. Incoherent compounding is a signal processing technique that involves recombination of echoes from different propagation paths. Incoherent compounding only considers the amplitude information from received signals. Therefore, the interpretation of received signals is simplified. A favourable application of this technique is crack sizing. Coherent processing, in which the amplitude and phase of received signals are considered, would be beneficial when enhanced resolution is necessary. Both techniques would be applicable to multi-layered structures commonly found within the aerospace industry. Experimentally, using a B-scan, incoherent compounding can be demonstrated by mirroring the direct image onto the reflected image(2). This paper will discuss incoherent compounding using a linear phased array probe with FRD capture as an innovative post-processing method. Furthermore, the results confirm the sizing and location of a combination of side-drilled holes that mimic defects.

Investigation of crack sizing using ultrasonic phased arrays with signal processing techniques

Conventional ultrasonic array technology has limitations associated with detection and sizing of defects, particularly for difficult structures such as the carbon fibre composites and titanium composites used widely for construction of aircraft components. This paper describes an investigation of data acquisition and signal processing techniques for crack sizing using electronically-scanned ultrasonic linear arrays. The data acquisition technique involves aperture number variance (that is, varying the number of active elements on the array) and the use of full raw data (FRD) capture. FRD capture accumulates the time domain responses from each transmitter-receiver pair manipulated on the array. With FRD capture, the raw data is collected from the array probe and reconstructed to produce a real-time image during the scan. However, the full data set is also stored and so beam steering and focusing become features of post-processing analysis. Experimental results obtained from an ultrasonic array inspection of difficult materials with and without FRD capture are compared for detection capability and precision of simulated crack sizing. A detailed description of the distributed data acquisition system has been described previously.

Publishing raw data and real time statistical analysis on e-journals

Authors of medical publications rarely provide their readers with the full raw data from their work but provide only the summarised statistical analysis. Indeed, publishing the raw data in a...

Good laboratory practices and other regulatory issues: A regulatory affairs view

AbstractCurrent international approaches differ regarding the definition of safety pharmacology studies, the numbers of these studies required for product support and whether it is necessary to conduct these studies in conformance with Good Laboratory Practices (GLPs). The European Union currently requires conformance with GLPs, Japan has a policy allowing full raw data audits of these studies, and the FDA does not require that these studies conform with GLPs unless they are performed in support of studies in the toxicology package (i.e. drug safety). Although the traditional position has been that it is impractical to conduct these studies in conformance with GLPs, reasons for maintaining this position are often incorrect and based on perception, rather than on understanding of the intent and extent of GLPs. Because it is highly possible that safety pharmacology studies may soon be required to conform with GLPs, this manuscript includes a review of GIP requirements, real, rather than perceived, impediments to GLP conformance, and practical suggestions for defining these types of specialized studies and performing them in such a way that they meet GLP requirments. © 1995 Wiley‐Liss, Inc.