4,473 research outputs found
DP-AdamBC: Your DP-Adam Is Actually DP-SGD (Unless You Apply Bias Correction)
No full textThe Adam optimizer is a popular choice in contemporary deep learning due to its strong empirical performance. However we observe that in privacy sensitive scenarios, the traditional use of Differential Privacy (DP) with the Adam optimizer leads to sub-optimal performance on several tasks. We find that this performance degradation is due to a DP bias in Adam's second moment estimator, introduced by the addition of independent noise in the gradient computation to enforce DP guarantees. This DP bias leads to a different scaling for low variance parameter updates, that is inconsistent with the behavior of non-private Adam, and Adam's sign descent interpretation. We propose the DP-AdamBC optimization algorithm, which corrects for the bias in the second moment estimation and retrieves the expected behaviour of Adam. Empirically, DP-AdamBC significantly improves the optimization performance of DP-Adam by up to 3.5% in final accuracy in image, text, and graph node classification tasks
A common HLA-DPA1 variant is associated with hepatitis B virus infection but fails to distinguish active from inactive Caucasian carriers
No full textBackground and Aims: Chronic infection with the hepatitis B virus (HBV) is a major health issue worldwide. Recently, single nucleotide polymorphisms (SNPs) within the human leukocyte antigen (HLA)-DP locus were identified to be associated with HBV infection in Asian populations. Most significant associations were observed for the A alleles of HLA-DPA1 rs3077 and HLA-DPB1 rs9277535, which conferred a decreased risk for HBV infection. We assessed the implications of these variants for HBV infection in Caucasians.
Methods: Two HLA-DP gene variants (rs3077 and rs9277535) were analyzed for associations with persistent HBV infection and with different clinical outcomes, i.e., inactive HBsAg carrier status versus progressive chronic HBV (CHB) infection in Caucasian patients (n = 201) and HBsAg negative controls (n = 235).
Results: The HLA-DPA1 rs3077 C allele was significantly associated with HBV infection (odds ratio, OR = 5.1, 95% confidence interval, CI: 1.9–13.7; p = 0.00093). However, no significant association was seen for rs3077 with progressive CHB infection versus inactive HBsAg carrier status (OR = 2.7, 95% CI: 0.6–11.1; p = 0.31). In contrast, HLA-DPB1 rs9277535 was not associated with HBV infection in Caucasians (OR = 0.8, 95% CI: 0.4–1.9; p = 1).
Conclusions: A highly significant association of HLA-DPA1 rs3077 with HBV infection was observed in Caucasians. However, as a differentiation between different clinical courses of HBV infection was not possible, knowledge of the HLA-DPA1 genotype cannot be translated into personalized anti-HBV therapy approaches
On the effects of non-linearities in DP systems
No full textIn offshore operations a trend is forming where vessels are more often required to do multiple short operations within a small-time frame. Traditional mooring systems require execution time far beyond the operation time. Dynamic positioning systems offer great advantages for short time span operations such as crew transfer or lift operations. Currently operations are planned based on DP capability plots and experience of captain and DPO. DP capability plots have little operational value as this is a static calculation and only provide information for average station keeping capability. During operations, the displacements made by the vessel around the DP set-point, also referred to as DP offset, are of great importance to determine the operability of an operation. Currently, the only way of calculating the DP offset is by conducting extensive time domain simulations, which are hard to integrate in the operational workflow of a DP vessel involved in walk-to-work operations. Therefore, a new approach is developed which predicts the vessel’s DP offset in the frequency domain, which enables a quick and robust calculation of the DP offset which is suited to merge into the on-board workflow. A frequency domain model is per definition a linear model. This leads to the main challenge of this research. A vessel operating on DP is non-linear. Currently there is no insight in what the effect is of non-linear components present in a DP system, on the linear approximation of a frequency domain model. To investigate the effect of non-linear components onto the DP frequency domain model, a time domain model is developed that is capable of systematically enabling/disabling different non-linear components. The time domain model will serve as the ’truth’ in this research as no actual vessel data is available. Furthermore, this helps identify the effects more easily, as the input for both models are identical. From the time domain model transfer functions can be derived that serve as the basis for the frequency domain model. The transfer function is a linear relation between two variables. In this case, between second order wave drift forces and displacement of the vessel in surge, sway and yaw direction. The following non-linear components are investigated in this research: Thruster ramp up, thruster turning rate, forbidden zones, saturation and thruster allocation. Thruster allocation is present in each model that will be tested, as this is an essential part of a DP system. Using two methods of determining transfer functions the model and the effects of all non-linear components are tested. The model is subjected to a variety sea-state, with different wave directions. Both methods offer similar results even though different approaches to determine the transfer functions are used. The selected method is capable of accurately predicting vessel offsets, although some extreme offsets are not captured. It is concluded that the presence of non-linear components have little to no effect on the DP offset as calculated by the time domain model. Because natural frequencies characteristic to these non-linear components are expected to exist at much higher frequencies that naturally present in second order wave drift forces. Thus, making a linear frequency domain model suitable for DP offset forecasting. It is advised to investigate the effect of including 2D input spectra as this is expected to improve the current model.<br/
Differentially Private GAN for Time Series
No full textGenerative Adversarial Networks (GANs) are a modern solution aiming to encourage public sharing of data, even if the data contains inherently private information, by generating synthetic data that looks like, but is not equal to, the data the GAN was trained on. However, GANs are prone to remembering samples from the training data, therefore additional care is needed to guarantee privacy. Differentially Private (DP) GANs offer a solution to this problem by protecting user privacy through a mathematical guarantee, achieved by adding carefully constructed noise at specific points in the training process. A state-of-the-art example of such a GAN is Gradient Sanitized Wasserstein GAN, (GS-WGAN), \cite{chen2021gswgan}. This model is shown to create higher quality synthetic images than other DP GANs. To extend the applicability of GS-WGAN we first reproduce and extend the evaluation, verifying that the model outperforms DP-CGAN by an average of 40\% when assessed across three qualitative metrics and two datasets. Secondly we propose improvements to the architecture and training procedure to make GS-WGAN applicable on timeseries data. The experimental results show that GS-WGAN is fit for generating synthetic timeseries through promising experimental results.[1] D. Chen, T. Orekondy, and M. Fritz, “Gs-wgan: A gradient-sanitized approach for learning differentially private generators,” 2021CSE3000 Research ProjectComputer Science and Engineerin
BNN-DP: Robustness Certification of Bayesian Neural Networks via Dynamic Programming
No full textIn this paper, we introduce BNN-DP, an efficient algorithmic framework for analysis of adversarial robustness of Bayesian Neural Networks (BNNs). Given a compact set of input points T ⊂ Rn, BNN-DP computes lower and upper bounds on the BNN's predictions for all the points in T. The framework is based on an interpretation of BNNs as stochastic dynamical systems, which enables the use of Dynamic Programming (DP) algorithms to bound the prediction range along the layers of the network. Specifically, the method uses bound propagation techniques and convex relaxations to derive a backward recursion procedure to over-approximate the prediction range of the BNN with piecewise affine functions. The algorithm is general and can handle both regression and classification tasks. On a set of experiments on various regression and classification tasks and BNN architectures, we show that BNN-DP outperforms state-of-the-art methods by up to four orders of magnitude in both tightness of the bounds and computational efficiency.Team Luca Laurent
The discriminative stimulus for punishment or S (Dp)
No full textdiscriminative stimulus for punishmentPresents the term discriminative stimulus for punishment and an accompanying symbol S-super(Dp). In describing a study of stimulus control in response-cost punishment with humans, the author notes that there was difficulty in writing a clear and succinct stimulus correlated with punishment conditions without using such a term. The author believes that the adoption of S-super(Dp ) as a discriminative stimulus for punishment would reserve S-super(D ) for reinforcement-only conditions, thus eliminating confusion as well as the need for additional descriptors. Furthermore, use of S-super(Dp ) would obviate the need for a term designating a stimulus correlated with the absence of punishment conditions
A Frequency Domain Approach to Estimate DP Footprint
No full textOffshore operations are increasingly executed by vessels operating on dynamic positioning (DP) due to advantages it has for short operations, such as maintenance and crew transfer but also for track keeping operations and projects in deeper water. Currently, DP capability plots are used to indicate whether an operation can be performed, but this is based on static calculations and it does not consider the offsets that are found due to motions of the vessel, while the offsets can be critical in case a structure is present or if a certain accuracy of a station keeping operation is necessary. Currently, to determine the offsets of a vessel on DP, the system is modelled using a time domain approach. Multiple simulation runs are carried out to calculate the expected extreme offsets in a given time interval, which is known as the most probable maximum offset (MPM offset). This is a complex and time consuming process and therefore it is not always done in practice before an operation starts. In this research, a method is developed to estimate the surge MPM offsets of a vessel on DP due to the wave drift forces using a frequency domain approach, as this can lead to faster estimates. But two problems are faced: The first is that there is no mathematical model available describing a vessel on DP which can be used to accurately calculate the offsets in the frequency domain. The second problem is that there is no known relation between the offsets and the extreme behaviour, which leads to the MPM offsets. To find the MPM offsets, first of all a one-dimensional time domain model of a vessel on DP is made, considering only the surge degree of freedom. Next, to determine the surge offset response using a frequency domain approach, the differential equations of the system are linearised, which gives the transfer functions from the environmental forces to the surge offsets. This is used to estimate the surge offset response in the frequency domain. The accuracy of this method is determined by comparing the root mean square value (RMS) and the zero up crossing period to that of the time domain results. Then, two alternative methods are developed to calculate the MPM offsets from the surge offset response characteristics directly. The methods use the RMS and zero up crossing period of the surge offsets calculated in the frequency domain, to determine the extreme behaviour statistically without the use of multiple simulation runs. Using the linearised model of the vessel on DP, it is found that the surge offset response can be calculated within an accuracy of 4% of the results generated by the time domain simulations, based on the root mean square value and the zero crossing period. Therefore, an accurate estimate of the surge offsets is found using a frequency domain approach. Both alternative MPM offset calculation methods, using the surge offset results of the linearised model, give MPM offset estimates within approximately 10% of the results of the time domain simulation approach. Taking into account the strong variability found in the MPM offsets calculated by the time domain approach, the estimates from the frequency domain approach are regarded as good estimates. It is concluded that the methods developed in this research can lead to faster and therefore timely MPM offset estimates to use for the safety and accuracy of operations. This can improve the way of working for many offshore operations where often no use is made of MPM offset estimates due to the disadvantages of the methods currently used in practice. It is recommended to extend the method developed in this research to three degrees of freedom, such that it can be implemented in practice.Offshore and Dredging Engineerin
Investigation of Unexpected Dynamic Vessel Response of a SSCV during DP-Operations
No full textDuring offshore construction with a Semi-Submersible Crane Vessel (SSCV) the vessel is station keeping by means of Dynamic Positioning (DP). DP enables a vessel to keep position and heading by utilizing its own propulsion, while being exposed to the environmental forces caused by waves, wind and current. In a comparison study between offshore measurements and time-domain simulations, an increased vessel/thruster response for the offshore measurement was observed. The results revealed, that in operational environmental conditions, the measured vessel response shows increased oscillations in Surge, Sway & Yaw, with periods of approx. 3-5 min. The goal of this thesis was to determine the causes for such an increased dynamic vessel response, which are not captured in time-domain simulations. At present, numerical methods and time-domain simulations that assess the DP performance of a vessel (e.g. aNySIM) assume a quasi-static current of which the variation is only caused by the tides. One thesis is, that time-varying currents on a scale of 1 to 5 minutes can cause an increased vessel response. Furthermore, the DP System as it is used onboard is not captured in aNySIM simulations. This means that the characteristics of the DP System onboard are but not represented in full detail in time-domain simulations. Vortex Induced Motion (VIM) has previously been observed to affect multi column floaters. However, the influence on a SSCV during DP-operations has not yet been studied. In this thesis it was investigated whether the unexpected increased motions originate from time-varying currents, VIM or the DP System itself. A significant challenge was posed to find current measurement data with a small enough time step to confirm the presence of such time-varying currents. One 45 min current measurement with a sampling rate of 1Hz became available. This measurement shows that time-varying currents on a scale of 1 to 5 minutes exist and cannot be assumed to be quasi-static. More research is required to confirm their presence. However, with this limited available data it was shown that time-varying currents can cause an unexpected motion response of a SSCV during DP-operations.In order to obtain the DP System characteristics, a DP Response Amplitude Operator (RAO) assessment was conducted wherein the spring and damping terms were derived as demanded by the controller and as experienced by the vessel. This assessment showed that between 20% to 67% of the demanded critical damping is lost over the control loop of the DP System. These damping losses cause the vessel to overshoot and the damped natural period to decrease. To determine the effects of VIM, current load tests carried out on the hull of an SSCV have been investigated. It was determined that strong currents on column type floaters can cause fluctuating forces and moments that originate from vortex shedding. Further, with real-time simulations it was demonstrated, that also VIM causes an increased, previously unknown, motion response of a SSCV. Lastly, a method was developed to extrapolate the current load test results to velocities below 2kn. Subsequently, in time-domain simulations it was shown that for current velocities above 0.5kn, the vessel experienced forces and moments that caused it to fluctuate around its setpoint
Beantwoording van vragen DP Rijnmond-Drechtsteden: Onderzoek faalkans - afsluitbare waterkeringen in de Rijnmond
No full textOp basis van twee eerdere onderzoeksresultaten wordt nader onderzoek gedaan naar effecten van klimaatverandering op de maatgevende hoogwaterstanden. Om tot een betere afstemming te komen zijn vijf nadere onderzoeksvragen gesteld door Delta-deelprogramma Rijnmond-Drechtsteden (DP-RD). Specifiek gaat het om de faalkans van de Maeslantkering.kennis voor klimaa
- …
