Datasets:

SouthernCrossAI
/

Arxiv

Dataset card Data Studio Files Files and versions Community

Split (1)

train · 132k rows

Filename stringlengths 22 64	Paragraph stringlengths 8 5.57k
Processed_Predicting_Melbourne_Ambulance_Demand_using_Kernel.txt	Much attention has been given to predicting the aggregate ambulance demand as a temporal process, using autoregressive moving average models (Channouf et al., 2007), factor models (Matteson et al., 2011), and spectral analysis (Vile et al., 2012). Few studies have modeled spatio-temporal am- bulance demand well. Setzler, Saydam and Park (2009) use artiﬁcial neural networks, but fail to improve over the industry method. A recent study by Zhou et al. (2015) predicts ambulance demand for Toronto, Canada using a time-varying Gaussian mixture model (GMM). This method is more accu- rate than the industry practice, but, as the authors point out, extending it to incorporate spatial boundaries would be prohibitively expensive. While it may not be essential for Toronto since the city is almost rectangular in shape, it becomes important for Melbourne; it is diﬃcult for ellipsoidal Gaussian components to model demand well on the highly complex spa- tial domain of Melbourne. Another study by Zhou and Matteson (2015) considers a spatio-temporal weighted kernel density estimation (KDE) to predict Toronto’s ambulance demand. It gives similarly accurate predictions as GMM (both much better than industry), showing promise for KDE.
Processed_Predicting_Melbourne_Ambulance_Demand_using_Kernel.txt	streams (Aggarwal, 2003). It allows for rapid identiﬁcation of “hotspots” and their evolutions in time and space. However, implementing a naive KDE is not satisfactory for our application. The chosen bandwidth necessarily has to be quite large given the data sparsity, smoothing inappropriately across boundary features and disregarding the underlying urban geography.
Processed_Predicting_Melbourne_Ambulance_Demand_using_Kernel.txt	Few studies have focused on modeling spatial or spatio-temporal point processes on complex spatial structures. Most studies assume a boundary deﬁned a priori (polygon or pixelated). If not, ad hoc methods based on the convex hull of all observed points are typically used (Ripley and Rasson, 1977). This invariably results in a convex boundary that may be inaccurate where data is sparse. Even with a boundary optimally deﬁned, few methods are equipped to handle complex boundary features. Ramsay (2002) proposes a ﬁnite window smoother with known boundary conditions computed using an expensive ﬁnite element approach. Building on that, Wood, Bravington and Hedley (2008) model the boundary condition as a loop of wire and the point process as a soap ﬁlm suspended from the boundary wire. They repre- sent this smoother as a penalized basis, compute it via multigrid, and select smoothness via generalized cross-validation. They acknowledge the lack of an elegant solution when the boundary conditions are unknown. Apart from boundary, other spatial characteristics, such as neighborhood structures and road networks, are rarely incorporated in modeling. We propose a method that can eﬃciently capture and exploit a wide range of spatial characteris- tics. We draw from theory and methods developed in manifold learning.
Processed_Predicting_Melbourne_Ambulance_Demand_using_Kernel.txt	Manifold learning, a branch of machine learning, is concerned with esti- mating and exploiting the underlying structures of data. The assumption is that data in a high-dimensional space resides on or near a lower-dimensional sub-manifold. In practice, we do not have access to this sub-manifold, but we can approximate it from a point cloud, i.e., a mass of historical data. The most common method is to construct an adjacency graph of this point cloud and make use of the properties and structures of this graph. This idea has led to many popular learning methods, including isomap (Tenebaum, de Silva and Langford, 2000), local linear embedding (Roweis and Saul, 2000), and Laplacian eigenmaps (Belkin and Niyogi, 2003). These meth- ods were initially designed for data representation or visualization, but have been adapted for semi-supervised classiﬁcation (Belkin and Niyogi, 2004), and clustering (Ng, Jordan and Weiss, 2001; Shi and Malik, 2000).
Processed_Predicting_Melbourne_Ambulance_Demand_using_Kernel.txt	cloud data, the method classiﬁes new examples by constructing kernels on the labeled data that warp to the geometry of the point cloud. This geom- etry is represented by the adjacency graph of the point cloud. Smoothing orthogonal to this geometry is penalized heavily, whereas smoothing along this geometry is not. This method is designed for high-dimensional classiﬁ- cation, and has good performance on text and image data.
Processed_Predicting_Melbourne_Ambulance_Demand_using_Kernel.txt	Drawing from this idea, we propose a novel method for modeling spatio- temporal point processes against complex spatial structures and features. To predict ambulance demand for a future time period, we have a sparse set of historical data that is very relevant for this prediction (labeled data). We ﬁt a KDE on them, but warp the kernels to a larger set of historical data regardless of their relevance to this predictive task (point cloud). This point cloud describes our belief about the spatial structure on which the labeled data lies. It captures exterior and interior boundaries without needing to explicitly deﬁne boundaries and boundary conditions. It also incorporates a wide range of complex spatial similarities and discontinuities, such as roads, city blocks, and neighborhoods of varying shapes and densities. Intuitively, this warping can be thought of as a regularization that penalizes radical departure from and encourages ﬂow of information along our intuition of the geography. In a Bayesian sense, it can also be thought of as imposing a prior based on how similar or diﬀerent the point process is across diﬀerent locations. Such a regularization or prior is especially beneﬁcial when the labeled data is sparse. We select the kernel bandwidth and the degree of warping eﬃciently via cross-validation. Both of these parameters can be made time- and/or location-speciﬁc.
Processed_Predicting_Melbourne_Ambulance_Demand_using_Kernel.txt	We implement this method on ambulance demand data from Melbourne in years 2011 and 2012. Altogether there are 696, 975 realized events. Each event contains the time and location that the ambulance was dispatched to. The proposed kernel warping model gives signiﬁcantly more accurate predictions than previous approaches, including the MEDIC method as an industry practice, unwarped KDE, and GMM.
Processed_Predicting_Melbourne_Ambulance_Demand_using_Kernel.txt	We develop the kernel warping model in Section 2. We construct an un- warped KDE in 2.1, warp the kernels to the point cloud in 2.2, and allow for time- and location-speciﬁc warping in 2.3 for the Melbourne data. Some details on computation are included in 2.4. We show the empirical results for predicting Melbourne ambulance demand in Section 3, and conclude in Section 4.
Processed_Predicting_Melbourne_Ambulance_Demand_using_Kernel.txt	intervals T = {1, 2, . . .}. Let st,i be the location of the i-th ambulance demand arising from the t-th time period, for i ∈ {1, . . . , nt}, where nt is the total number of ambulances demanded in the t-th period. Since a non-homogeneous Poisson process (NHPP) is a natural model for spatial point process (Diggle, 2003; Møller and Waagepetersen, 2004), we assume {st,i : i = 1, . . . nt} for each time period t independently follow an NHPP over S, with positive intensity function λt. We decompose the intensity function as λt(s) = δtft(s), for s ∈ S. Here, δt = (cid:82) S λt(s) ds is the aggregate demand intensity over the spatial domain, and ft(·) is the continuous spatial density of the demand at time t such that ft(s) > 0 and (cid:82) S ft(s)ds = 1. Therefore, iid∼ ft(·) for i ∈ {1, . . . , nt}. The for each t, nt\|λt ∼ Poisson(δt) and st,i\|λt, nt usual practice is to model {δt} and {ft} separately. As mentioned before, numerous studies have proposed sophisticated and accurate methods for estimating {δt}. We thus focus on predicting the spatio-temporal demand density {ft}, which is more challenging and less studied.
Processed_Predicting_Melbourne_Ambulance_Demand_using_Kernel.txt	density varies substantially in space (downtown vs. neighborhoods) and time (midnight vs. rush hours); we may be motivated to consider a spatial- and/or time-varying H.
Processed_Predicting_Melbourne_Ambulance_Demand_using_Kernel.txt	In our application, we randomly sample 1000 historical observations as the point cloud for each “component” (to be explained in Section 2.3). We denote the set of point cloud data as {zi} for i ∈ {1, . . . , Z}. See Figure 2 (a) for an example cloud of 1000 points over the entire city of Melbourne.
Processed_Predicting_Melbourne_Ambulance_Demand_using_Kernel.txt	For our application, we ﬁnd that predictive accuracy is not sensitive to the random sampling of the point cloud data. If it were, a larger point cloud might be needed, or predictions might be repeated and averaged over several point cloud samples.
Processed_Predicting_Melbourne_Ambulance_Demand_using_Kernel.txt	Step 2 [Constructing the adjacency graph]: We construct a graph with nodes at each point in the point cloud and edges connecting points that are close. We represent this graph using a symmetric, positive semideﬁnite adjacency matrix A.
Processed_Predicting_Melbourne_Ambulance_Demand_using_Kernel.txt	In the Melbourne application, we use n = 5 nearest neighbors and binary weights to construct A. Figure 2 (a) shows the adjacency graph of a sample point cloud of size 1000. Again, we ﬁnd our predictive accuracy to be insen- sitive towards any reasonable variations in these choices.
Processed_Predicting_Melbourne_Ambulance_Demand_using_Kernel.txt	block diagonal matrix, where each block is the respective Laplacian matrix for each connected component.
Processed_Predicting_Melbourne_Ambulance_Demand_using_Kernel.txt	Here is the intuition of the Laplacian matrix. The (discrete) point cloud adjacency graph is an empirical approximation to our target (continuous) manifold of Melbourne geography. The (discrete) graph Laplacian matrix L is then an approximation to the (continuous) Laplace-Beltrami operator on this manifold. The Laplace-Beltrami operator is a manifold generalization of the Laplace operator, which is a linear second order diﬀerential operator on functions (in our case, kernels). This L induces a semi-norm on kernels which penalizes changes between adjacent nodes. There is a close analogy to heat ﬂow; the heat (partial diﬀerential) equation has a Laplace operator in space. Intuitively, L guides how information (heat) spreads on the spatial structure (manifold approximated by graph) from any initial KDE (initial heat distribution).
Processed_Predicting_Melbourne_Ambulance_Demand_using_Kernel.txt	in which kx = [k(x, z1 \| H), . . . , k(x, zZ \| H)] and ks = [k(s, z1 \| H), . . . , k(s, zZ \| H)] are vectors of kernels evaluated at x or s and the point cloud data {zi}. Matrix K = [k(zi, zj \| H)]i,j∈{1,...,Z} is a symmetric matrix of kernels evaluated at all pairs of point cloud data, and I is a Z by Z identity matrix. The parameter λ > 0 represents the degree of deformation. When λ = 0, we have ˜k = k. When λ → ∞, ˜k approaches a positive constant on the point cloud (steady state heat distribution).
Processed_Predicting_Melbourne_Ambulance_Demand_using_Kernel.txt	Equation (2) is obtained by warping the Reproducing Kernel Hilbert Space (RKHS) associated with the chosen kernel. We modify the RKHS with a point-cloud semi-norm λL. This deforms the kernel k along a ﬁnite- dimensional subspace given by the point cloud data. The modiﬁed RKHS is shown to be another RKHS, i.e., ˜k is a properly deﬁned kernel. See Sind- hwani, Niyogi and Belkin (2005) and Belkin, Niyogi and Sindhwani (2006) for more details (they use the point cloud semi-norm of λLp; we consider the simpliﬁed case where p = 1).
Processed_Predicting_Melbourne_Ambulance_Demand_using_Kernel.txt	violation of the spatial structure implied by the point cloud are penalized. Lastly, in the Bayesian framework, kernel warping can informally be thought of as imposing a data-dependent informative prior to describe our belief of the data geometry.
Processed_Predicting_Melbourne_Ambulance_Demand_using_Kernel.txt	We replace the regular Gaussian kernel k in Equation (1) with the new warped kernel ˜k deﬁned in (2) to predict the density of ambulance demand fu at a future time period u. We set a priori the sliding window width M , the point cloud data type / size, the number of nearest neighbors n, and the weights used to construct the Laplacian matrix. We estimate the Gaussian kernel bandwidth H and the degree of deformation λ.
Processed_Predicting_Melbourne_Ambulance_Demand_using_Kernel.txt	We show in Figure 2 (b) and (c) examples of warping kernels. Three kernels of bandwidth H = diag(2, 2) are placed on three observations circled in red in Figure 2 (a). They are warped towards the point cloud in (a) with degree of deformation λ = 0.5 (b) and 2 (c). With a larger λ, the kernels conform to the spatial boundary and characteristics to a greater extent.
Processed_Predicting_Melbourne_Ambulance_Demand_using_Kernel.txt	Fig 2: Examples of kernel warping: (a) the adjacency graph of a sample point cloud of size 1000; three observations are highlighted in red; (b) and (c), warped kernels centered at the these three observations with degrees of deformation λ = 0.5 and 2, respectively.
Processed_Predicting_Melbourne_Ambulance_Demand_using_Kernel.txt	periods. For each hour, we further discretize the spatial domain into a small number of regions, as motivated by the behavior of labeled data for that time period. We call each subregion of each hour a component, and perform estimations and predictions independently on each component. The spatial discretization splits a global point cloud into local ones, cuts all edges con- necting across regions, and decomposes the Laplacian matrix into blocks. Labeled data are also matched into components. We estimate a separate set of H and λ for each component by cross-validation (details in Section 2.4). We discretize spatially by clustering. For any given future time period, we cluster on its labeled data (about 300 observations). We allow diﬀerent num- bers of clusters and clustering conﬁgurations for each time period. In our application, this gives more accurate predictions than imposing a universal clustering conﬁguration across time. We also obtain better results by cluster- ing on labeled data rather than clustering on the point cloud data (the point cloud is much more similar across time than the labeled data). In the case of Melbourne, spatial characteristics across time are diﬀerent enough that the gain in personalized modeling exceeds the loss in stablization oﬀered by a common arrangement.
Processed_Predicting_Melbourne_Ambulance_Demand_using_Kernel.txt	We choose to cluster using K-means based on Euclidean distance. K- means is fast, clusters all points, and gives even clusters. Even cluster sizes are desirable because a very small cluster does not provide enough labeled data to reliably estimate parameters via cross-validation. To avoid this, we set a threshold minimum number of points in any cluster. We set the thresh- old at 15 points, which in practice limits the number of clusters to be be- low 8. If we fail to clear this threshold, we lower the number of clusters. Density-based clustering algorithms such as DBSCAN (Ester et al., 1996) and shared nearest neighbors (Ert¨oz, Steinbach and Kumar, 2003) do not classify all points, and do not allow easy speciﬁcation of the number of clus- ters. Graph-based clustering algorithms such as aﬃnity propagation (Frey and Dueck, 2007) and spectral clustering (Ng, Jordan and Weiss, 2001) do not cluster on Euclidean distance, and may be less intuitive for spatial point patterns. In our case, hierarchical clustering gives very uneven cluster sizes. For each time period, we binary search for the best number of clusters based on validation likelihood. Increasing the number of clusters leads to, on the one hand, an additional 1000 points to the cloud and the ﬂexibility to customize parameters locally, but on the other hand, sparser labeled data for each cluster and reduced stability in parameter estimation. It is an empirical question for each time period whether we have enough labeled data to aﬀord this increase in complexity. For Melbourne, we ﬁnd the number of clusters to be largely proportional to the size of labeled data.
Processed_Predicting_Melbourne_Ambulance_Demand_using_Kernel.txt	We choose H and λ for each component using 5-fold cross-validation to maximize average validation likelihood. We implement a surrogate, derivative- free optimization procedure called the stochastic radial basis function (RBF) method (Regis and Shoemaker, 2007, 2009). It is a fast algorithm for global optimization of computationally expensive objective functions. Each itera- tion builds an RBF model to approximate the expensive function, selects subsequent candidate points, and evaluates them in parallel. We choose this approach because our objective function (likelihood) evaluation is not in- stantaneous. It takes between 0.5 and 4 seconds, depending on the sizes of the labeled data and point cloud (Python code on a personal computer). We also do not have simple derivative computations. In our experience, 100 such evaluations are suﬃcient to provide a good optimum, competitive to those found by grid search, pattern search, or evolutionary algorithms. However, a wide range of optimization tools can be applied here.
Processed_Predicting_Melbourne_Ambulance_Demand_using_Kernel.txt	In our application, we ﬁnd a typical optimal α to be between 0.05 and 0.3. We need a concentration of heat which is then spread or warped to the point cloud. A typical optimal λ is between 0 and 2. Most time periods choose between 1 and 3 spatial components. We warm start the binary search for the number of clusters based on the size of labeled data. The best conﬁguration is usually found within 3 searches.
Processed_Predicting_Melbourne_Ambulance_Demand_using_Kernel.txt	Given the prominent weekly seasonality, we believe that the corresponding parameter values are also similar from week to week. In fact, we believe that the nature of deformation and smoothing does not vary signiﬁcantly over several months, and thus only estimate the parameters for a one-week cycle once every few months. With the most recent weekly set of parameter values, we predict forward in an online fashion with a sliding window of M = 8 weeks, making use of the most recent 8 weeks of data available. Each prediction is instantaneous.
Processed_Predicting_Melbourne_Ambulance_Demand_using_Kernel.txt	ally more eﬃcient than one massive global point cloud. There are ways to optimize this computation, including using right division instead of inver- sion, saving pre-computed kernel evaluation matrices and vectors, exploiting sparse, banded-diagonal Laplacian matrix, using a tree-based algorithm for fast KDE computation (Gray and Moore, 2003), and using a look-up table for Gaussian densities (most of these optimizations are not used in our im- plementation). The computation is “embarrassingly parallelizable”, across validation likelihood evaluations and across spatio-temporal components.
Processed_Predicting_Melbourne_Ambulance_Demand_using_Kernel.txt	Fig 3: Log predictive densities using spatio-temporal kernel warping for March 2, 2011 (Wednesday) at (a) 2 - 3 am (night), and (b) 2 - 3 pm (day). For time period (a), we have sparse data and cross-validate to choose 1 spatial component. For time period (b), we have more data and choose 5 spatial components.
Processed_Predicting_Melbourne_Ambulance_Demand_using_Kernel.txt	Figure 5 shows the log predictive density using the MEDIC method, un- warped KDE (PI), and GMM (30 components) for March 2, 2011 at 2 - 3 pm. These densities are comparable with Figure 3 (b), which shows the log predictive density for the same period predicted by the proposed kernel warping. Even with 400 labeled data, The MEDIC method gives exceedingly noisy predictions, while unwarped KDE and GMM produce over-smoothed densities that do not adapt well to the spatial features of Melbourne.
Processed_Predicting_Melbourne_Ambulance_Demand_using_Kernel.txt	Fig 4: Log predictive densities for March 2, 2011 (Wednesday) at 10 - 11 am using (a) spatio-temporal kernel warping (3 spatial clusters), and (b) temporal kernel warping. The density in (a) shows slightly more details.
Processed_Predicting_Melbourne_Ambulance_Demand_using_Kernel.txt	in which {˜su,i} are the test data, and ˆfu(·) is the predictive density for period u obtained by various methods. For the MEDIC method, we normalize cell counts to discrete density by dividing over the total count in each period.
Processed_Predicting_Melbourne_Ambulance_Demand_using_Kernel.txt	where C is the number of 1 km2 cells in the rectangular observation window, yu,c and ˆyu,c are the actual and predicted count for period u and cell c respectively. For RMSEB, we use cells c within the pixelated boundary B and C as the number of 1 km2 cells within this boundary.
Processed_Predicting_Melbourne_Ambulance_Demand_using_Kernel.txt	Fig 5: Log predictive densities using comparison methods for 2 - 3 pm on March 2, 2011 (Wednesday): (a) the MEDIC method (an industry practice); (b) unwarped KDE with bandwidth selected by the plug-in method (PI); (c) time-varying Gaussian mixture model with 30 components. These densities are to be compared to Figure 3 (b), which is the prediction using kernel warping for the same period.
Processed_Predicting_Melbourne_Ambulance_Demand_using_Kernel.txt	the variations of some of these metrics across time periods. Kernel warping has not only the best mean performance, but also the smallest variations across time periods.
Processed_Predicting_Melbourne_Ambulance_Demand_using_Kernel.txt	Table 1 Mean predictive accuracies across all 1-hour periods in March 2011 of the proposed kernel warping and competing methods. Kernel warping outperforms the competing methods.
Processed_Predicting_Melbourne_Ambulance_Demand_using_Kernel.txt	Fig 6: Box-plots of predictive accuracies of kernel warping (S-T parameters), GMM (30 comp), KDE (PI bandwidth), and the MEDIC method (an indus- try practice) over 672 test periods, as measured by average log score (left, less negative is better), RMSEB (middle, smaller is better), and ANSCB (right, smaller is better).
Processed_Predicting_Melbourne_Ambulance_Demand_using_Kernel.txt	resolution modeling. Most current industry practices and earlier studies are ill-equipped to address these challenges simultaneously. We propose a kernel warping method that smooths intelligently towards geographical characteris- tics. We demonstrate that our proposed method predicts ambulance demand in Melbourne more accurately than the state of the art in the practice and research of ambulance demand prediction.
Processed_Predicting_Melbourne_Ambulance_Demand_using_Kernel.txt	To predict ambulance demand for any hour, we use a spatio-temporal kernel density estimator on the sparse set of most similar labeled data, but warp these kernels to a larger point cloud drawn from all historical observations regardless of labels. We construct an adjacency graph on this point cloud to approximate Melbourne’s spatial boundaries, neighborhoods, and road networks in a data-driven manner. Kernels on labeled data are warped to encourage ﬂow along and penalize ﬂow orthogonal to this graph. Kernel warping circumvents the need to deﬁne boundaries and bound- ary conditions, which are often diﬃcult in the practice of modeling point patterns on complex spatial domains. It also captures and exploits ﬁner- grain internal spatial structures other than boundary features, which can be prominent in various heterogeneous environments such as cities, buildings, mountains, and forests. Kernel warping is not limited to density estimation. It can be adapted to model a wide range of functions and surfaces. It can be used to perform a broad set of tasks including prediction, classiﬁcation, clustering, and visualization. Inferences on uncertainty, if desired, can be ob- tained by assessing cross-validation variance and warping kernels to diﬀerent samples of point clouds. There is much ﬂexibility in designing the point cloud and its Laplacian. We oﬀer some discussions on these in the context of spa- tial and spatio-temporal point patterns. We also oﬀer eﬃcient estimation of kernel bandwidth and degree of warping local to time periods and locations via cross-validation. The proposed method is straightforward to implement and easy to experiment with. The tools we have developed can be easily gen- eralized to model a wide range of spatial or spatio-temporal point process on complex spatial domains.
Processed_Predicting_Melbourne_Ambulance_Demand_using_Kernel.txt	Acknowledgements. The authors sincerely thank Ambulance Victoria for sharing their data. We also thank the support from a Xerox Faculty Research Award and NSF grant DMS-1455172.
Processed_Predicting_Melbourne_Ambulance_Demand_using_Kernel.txt	Anscombe, F. J. (1953). Contribution of discussion paper by H. Hotelling ‘New light on the correlation coeﬃcient and its transforms’. Journal of the Royal Statistical Society: Series B 15 229 - 230.
Processed_Predicting_Melbourne_Ambulance_Demand_using_Kernel.txt	Belkin, M., Niyogi, P. and Sindhwani, V. (2006). Manifold regularization: a geometric framework for learning from labeled and unlabeled examples. The Journal of Machine Learning 7 2399 - 2434.
Processed_Predicting_Melbourne_Ambulance_Demand_using_Kernel.txt	Channouf, N., L’Ecuyer, P., Ingolfsson, A. and Avramidis, A. (2007). The applica- tion of forecasting techniques to modeling emergency medical system calls in Calgary, Alberta. Health Care Management Science 10 25–45.
Processed_Predicting_Melbourne_Ambulance_Demand_using_Kernel.txt	Duong, T. and Hazelton, M. L. (2005). Cross-validation bandwidth matrices for mul- tivariate kernel density estimation. Scandinavian Journal of Statistics 32 485–506. Ert¨oz, L., Steinbach, M. and Kumar, V. (2003). Finding clusters of diﬀerent sizes, shapes and densities in noisy, high dimensional data. In Proceedings of the SIAM In- ternational Conference on Data Mining 47 - 58.
Processed_Predicting_Melbourne_Ambulance_Demand_using_Kernel.txt	Ester, M., Kriegel, H. P., Sander, J. and Xu, X. (1996). A density-based algorithm for discovering clusters in large spatial databases with noice. In ACM SIGKDD Inter- national Conference on Knowledge Discovery and Data Mining 226-231.
Processed_Predicting_Melbourne_Ambulance_Demand_using_Kernel.txt	Gray, A. G. and Moore, A. W. (2003). Nonparametric density estimation: toward computational tractability. In Proceedings of the SIAM International Conference on Data Mining.
Processed_Predicting_Melbourne_Ambulance_Demand_using_Kernel.txt	Google Maps (2015). Map of Melbourne, Australia. Web. Matteson, D. S., McLean, M. W., Woodard, D. B. and Henderson, S. G. (2011). Forecasting emergency medical service call arrival rates. Annals of Applied Statistics 5 1379–1406.
Processed_Predicting_Melbourne_Ambulance_Demand_using_Kernel.txt	Nakaya, T. and Yano, K. (2010). Visualising crime clusters in a space-time cube: an exploratory data analysis approach using space-time kernel density estimation and scan statistics. Transactions in GIS 14 223 - 239.
Processed_Predicting_Melbourne_Ambulance_Demand_using_Kernel.txt	Sindhwani, V., Niyogi, P. and Belkin, M. (2005). Beyond the point cloud: from trans- ductive to semi-supervised learning. In Proceedings of the 22nd International Conference on Machine Learning 824 - 831.
Processed_Predicting_Melbourne_Ambulance_Demand_using_Kernel.txt	Vile, J. L., Gillard, J. W., Harper, P. R. and Knight, V. A. (2012). Predicting am- bulance demand using singular spectrum analysis. Journal of the Operations Research Society 63 1556–1565.
Processed_Predicting_Melbourne_Ambulance_Demand_using_Kernel.txt	Zhang, Z., Chen, D., Liu, W., Racine, J. S., Ong, S. H., Cheng, Y., Zhao, G. and Jiang, Q. (2011). Nonparametric evaluation of dynamic disease risk: a spatio-temporal kernel approach. PLoS ONE 6.
Processed_Predicting_Melbourne_Ambulance_Demand_using_Kernel.txt	Zhou, Z. and Matteson, D. S. (2015). Predicting ambulance demand: a spatio-temporal kernel approach. In ACM SIGKDD International Conference on Knowledge Discovery and Data Mining to appear.
Processed_Predicting_Melbourne_Ambulance_Demand_using_Kernel.txt	Zhou, Z., Matteson, D. S., Woodard, D. B., Henderson, S. G. and Micheas, A. C. (2015). A spatio-temporal point process model for ambulance demand. Journal of the American Statistical Association 110 6–15.
Processed_Flux_Jumping_and_a_Bulk-to-Granular_Transition_in_.txt	In this letter, we report the results of field (H) and temperature (T) dependent magnetization (M) measurements of a pellet of uniform, large-grain sintered MgB2. We show that at low temperatures the size of the pellet and its critical current density, Jc(H) -- i.e. its M(H) -- ensure low field flux jumping, which of course ceases when M(H) drops below a critical value. With further increase of H and T the individual grains decouple and the M(H) loops drop to lower lying branches, unresolved in the usual full M(H) representation. After taking into account the sample size and grain size, respectively, the bulk sample and the grains were deduced to exhibit the same magnetically determined Jc s (e.g. 105 A/cm2, 20 K, 0T) and hence that for each temperature of measurement Jc(H) decreased monotonically with H over the entire field range, except for a gap within the grain-decoupling zone.
Processed_Flux_Jumping_and_a_Bulk-to-Granular_Transition_in_.txt	The recent discovery of intermediate-temperature superconductivity (ITC) in MgB2 by Akimitsu et al. 1 and its almost simultaneous explanation in terms of a hole-carrier-based pairing mechanism by Hirsch2, 3 has triggered an avalanche of studies of its structural, magnetic and transport properties4,5,6,7,8. In spite of the significant advancement made on critical current density of HTS conductors the best transport Jc of Ag/Bi2223 tapes is still limited by the grain connectivity. The study on the intergranular and intragranular critical current densities revealed that the former is smaller than the latter by at least one order of magnitude for the state-the-art Ag/Bi2223 tapes9. This characteristic feature of HTS conductors is clearly evident by the rapid drop of Jc in low magnetic fields which is defined as weak link regime. Preliminary results on the newly discovered 40 K superconductor, MgB2 reported by Larbalestier et al have indicated that Jc of this material is not limited by grain boundary weak links as HTS6. The high critical current density of 105 A/cm2 at 11K and low field and 2· 104 A/cm2 at 20K and 1T have been reported on the bulk MgB2 samples using magnetisation measurements5,6. As a further contribution to the field, we design a synthesis method to produce bulk MgB2 samples with large and uniform grain size. The results of magnetisation measurements performed on these samples allow us to compare and evaluate the intergranular and intragranular critical currents in MgB2.
Processed_Flux_Jumping_and_a_Bulk-to-Granular_Transition_in_.txt	Phase purity was determined by XRD and grain sizes by SEM. Sintering time and temperature determine the phase content. The XRD pattern for a sample sintered for 4 h/800 o C is shown in Fig. 1, which shows mostly MgB2 with only a small level of MgO contamination. The SEM results for the same sample are illustrated in Fig. 2. The grain size is of the order of 200 m m. It is likely that the large size grains were formed by a partially melting process. The sample appears to be dense, hard, and strong.
Processed_Flux_Jumping_and_a_Bulk-to-Granular_Transition_in_.txt	The transition temperature Tc was determined by using AC susceptibility in an AC field of amplitude 1 Oe and frequency 117 Hz. The results, depicted in Fig.3, indicate a Tc of 38.2 K with a very sharp transition, whose width of only 2K again confirms the high phase purity of the sample.
Processed_Flux_Jumping_and_a_Bulk-to-Granular_Transition_in_.txt	Where b T3 represents the low temperature specific heat and D M0(T) the magnetization temperature dependence in the form D M/(dD M/dT). The lack of specific heat data prevents a direct prediction of Mmax. However a back calculation using the experimental Mmax predicts a specific heat for MgB2 at 5 K of 3x10-6 J/cm3.
Processed_Flux_Jumping_and_a_Bulk-to-Granular_Transition_in_.txt	Evidence to be presented below indicates a sample, initially a bulk superconductor, that breaks down into a granular assembly beyond a certain critical value of field and temperature. Such a breakdown will also occur during a flux jump. The flux jumps depicted are partial, rather than complete. This suggests that a jump is terminated when the magnetization of the outer shell grains, and that of the residual core, drop below the critical value. Within the flux jump zone of Fig.4 (for T<10K and H<1T) we note the magnetization dropping to about one- fifth of its initial (virtual) value during the partial jump.
Processed_Flux_Jumping_and_a_Bulk-to-Granular_Transition_in_.txt	A magnetic critical current density can be derived from the height of the magnetization loop D M using a suitable variant of the "semi-Bean" relationship D M = k Jcd, where k is a constant and d the thickness that the sample presents to the applied field. D M itself versus H is plotted semi-logarithmically in Fig. 5. Here we note what are essentially a high-field and low-field regions of curves, the transition between which is complete at a fairly constant value of D M. Since the low field family covers three orders of magnitude in D M, the low-field sets are not visible in the usual linear type of M(H) loop (Fig.4).
Processed_Flux_Jumping_and_a_Bulk-to-Granular_Transition_in_.txt	The transition from low field step to high field step corresponds to a drop in D M by more than an order of magnitude. It is interesting to note that from the image of SEM of the sample the average grain size is less than one tenth of the sample size. It is evident that the drop in D M is attributable to the breakdown of grain matrix as a result of flux penetration to the grain boundaries which may contain impurities. In the low field region the current circulates mainly over the entire sample size, which is the intergranular current. The induced current circulates only in the individual grains in the high field region, which is the intragranular current [9]. The transition from low field step to high field step represents a transition from intergranular current regime to intragranular current regime.
Processed_Flux_Jumping_and_a_Bulk-to-Granular_Transition_in_.txt	Based on the full sample size at low fields and the grain size at high fields, the magnetic Jc is calculated using the relationship for a plate in perpendicular field: Jc = 20D M/(a-a2/3b). For low fields, the lateral dimensions of the sample are used: a=0.25cm, b=0.35cm, whereas for the high fields a=170 m m, b=260 m m. No calculation is applied to the intermediate-field transitional zone wherein the grain decoupling is taking place. The results are depicted in Fig. 6 wherein we note a continuous Jc field dependence throughout the entire field range.
Processed_Flux_Jumping_and_a_Bulk-to-Granular_Transition_in_.txt	In summary, we presented evidence for decoupling of the grains in sintered MgB2, both through partial flux jumping and a step in field-dependence of Jc. Similar step was also observed in the high-temperature superconductors. However a significant difference here is that the step occurs at much higher fields and it corresponds to a negligibly small drop of Jc (Fig. 5). Therefore, grain connections in MgB2 can sustain rather large Jc, in contrast to high- temperature superconductors. More significantly, the field dependence of inter-grain Jc seems to be the same as that of intra-grain Jc (Fig.6). The flux jump and step in Jc(H) was observed thanks to large value of Jc and large size of the grains. Our samples gave Jc of 105 A/cm2 at 20 K and zero field, as obtained from magnetic hysteresis loops with sweep rate of field of 50 Oe/s. The value of Jc would be much larger at lower temperature. However it was not possible to obtain it there from the hysteresis loop because of the flux jump.
Processed_Flux_Jumping_and_a_Bulk-to-Granular_Transition_in_.txt	Figure 1: XRD pattern for the sample showing that all the peaks can be indexed according to the published crystallography data of hexagonal MgB2 except for the small peaks of MgO as impurity phase.
Processed_Flux_Jumping_and_a_Bulk-to-Granular_Transition_in_.txt	Figure 2: SEM images showing that the sample consists of large and uniform gains with size ranging from 0.15mm to 0.3mm.
Processed_Flux_Jumping_and_a_Bulk-to-Granular_Transition_in_.txt	Figure 3: AC susceptibility, measured with frequency and amplitude of ac field of 117Hz and 1 Oe, respectively.
Processed_Flux_Jumping_and_a_Bulk-to-Granular_Transition_in_.txt	Figure 5: Width of the hysteresis loops, D M, as a function of field, with the sweep rate of the field 50Oe/s.
Processed_Flux_Jumping_and_a_Bulk-to-Granular_Transition_in_.txt	Figure 6: Critical current density vs. field, with corrections for different screening length of the currents in small and high fields.
Processed_IDOL:_A_Framework_for_IMU-DVS_Odometry_using_Lines.txt	an this optimization-based framework for IMU-DVS Odometry using Lines. Event cameras, also called Dynamic Vision Sensors (DVSs), generate highly asynchronous streams of events triggered upon illumination changes for each individual pixel. This novel paradigm presents advantages in low illumination conditions and high-speed motions. Nonetheless, this unconventional sensing modality brings new challenges to perform scene reconstruction or motion estimation. The proposed method offers to leverage a continuous-time representation of the inertial readings to associate each event with timely accurate inertial data. The method’s front-end extracts event clusters that belong to line segments in the environment whereas the back-end estimates the system’s trajectory alongside the lines’ 3D position by minimizing point-to-line distances between individual events and the lines’ projection in the image space. A novel attraction/repulsion mechanism is presented to accurately estimate the lines’ extremities, avoiding their explicit detection in the event data. The proposed method is benchmarked against a state-of-the-art frame-based visual-inertial odometry framework using public datasets. The results show that IDOL performs at the same order of magnitude on most datasets and even shows better orientation estimates. These ﬁndings can have a great impact on new algorithms for DVS.
Processed_IDOL:_A_Framework_for_IMU-DVS_Odometry_using_Lines.txt	Fig. 1: The proposed method, IDOL, estimates the system’s ego- motion based on the segmentation and position estimation of line segments in the environment.
Processed_IDOL:_A_Framework_for_IMU-DVS_Odometry_using_Lines.txt	of illumination. Dynamic Vision Sensors (DVSs), also called event-based cameras, are new sensor types that have a huge potential in addressing aforementioned limitations due to their extremely high temporal resolution and their HDR operative range [9], [10]. In contrast to traditional frame-based cameras, which periodically output intensity values for all pixels, these sensors output event tuples e = {t, x, y, p}, where t is the timestamp of the event, x, y are the image coordinates, and p is the event polarity. Such an event is triggered only in case a pixel’s intensity change is larger than a threshold, with p being the direction of that change. Consequently, in a static scene, events are only generated when the camera is in motion. The events form an asynchronous stream of data that provides reliable information even in the presence of fast motion. As of today, most developments in the event- based literature focused on techniques to aggregate the event data into key-frames in order to apply or adapt conventional frame-based VO algorithms. To fully leverage the potential of this novel type of modality, new algorithms and ego-motion estimation paradigms need to be developed.
Processed_IDOL:_A_Framework_for_IMU-DVS_Odometry_using_Lines.txt	such as lines, making point landmarks not necessarily the best representation for visual tracking in all the scenarios. In its current state, our front-end does not offer enough robustness to address the boxes scenarios of [11] because of their very high level of texture in the scene.
Processed_IDOL:_A_Framework_for_IMU-DVS_Odometry_using_Lines.txt	In this paper, we investigate the potential of directly using asynchronous events without any frame-like accumulation, for ego-motion estimation. A key element is the use of a continuous-time representation of the inertial data to con- strain the system pose at any time without relying on any mo- tion model. By tightly coupling IMU and event data together via the generation of inertial information at each event’s timestamp, a visual-inertial odometry (VIO) formulation that addresses and leverages the data asynchronism rigorously is achieved. Furthermore, instead of using traditional point features, we represent the environment using line segments. We introduce a new pipeline, IMU-DVS Odometry with Lines (IDOL), that detects line features in the event data and performs VIO by individually considering asynchronous events in a batch-optimization framework.
Processed_IDOL:_A_Framework_for_IMU-DVS_Odometry_using_Lines.txt	The remainder of the paper is organized as follows: In Section II, a summary of related work in event-based motion estimation, continuous state and measurement rep- resentation is presented. Section III provides an overview of the proposed framework IDOL while Section IV and Section V gives more details about the front-end and back- end, respectively. In Section VI, IDOL is evaluated on public datasets and compared to state-of-the-art in traditional frame- based VIO. Finally, Section VII provides a conclusion with an outlook on future work.
Processed_IDOL:_A_Framework_for_IMU-DVS_Odometry_using_Lines.txt	The major differences between the traditional frame-based and event-based vision make the latter an especially ap- pealing sensing modality for the task of VO in challenging scenarios where the performance of traditional imagers is compromised such as in HDR scenes or under fast camera motions. Despite the relatively recent interest of the commu- nity in event cameras, we can already identify several works on VO employing event cameras over the last years, starting from the ﬁrst 2D SLAM approach by Weikersdorfer et al. [12]. Years later, Kim et al. [13] achieved the estimation of the 6-degree of freedom (DoF) of the camera pose on generic 3D scenes using probabilistic ﬁltering. Rebecq et al. [14] proposed a parallel mapping and tracking approach that iteratively co-localize the pose of the camera against a local map of edges represented by a voxel grid.
Processed_IDOL:_A_Framework_for_IMU-DVS_Odometry_using_Lines.txt	Aforementioned approaches avoid the explicit deﬁnition of features at expenses of being relatively demanding in terms of computational resources. Modern event-based VIO pipelines such as [15]–[17], however, rely on detection and tracking of corners employing intermediate image-like representations from the accumulation of events. These ap- proaches make use of IMUs, proﬁting from their high-rate of inertial measurements and making them an appealing type of sensor to be combined with event cameras.
Processed_IDOL:_A_Framework_for_IMU-DVS_Odometry_using_Lines.txt	While the bulk of event-based VO approaches still rely on the traditional concept of key-frames, it is only natural that continuous-time approaches would emerge. One of the seminal works is introduced by Mueggler et al. [18], in which the state of the camera is estimated associating events with line segments evaluated both in simulation and real-world experiments using ﬁducial markers detected using intensity images. The same authors later expand their approach to consider inertial measurements in [19].
Processed_IDOL:_A_Framework_for_IMU-DVS_Odometry_using_Lines.txt	Over the years, signiﬁcant efforts in the community have been dedicated to the deﬁnition of reliable visual features for event data. It has led to the development of mainly corner detection and tracking approaches (e.g. [20]–[23]), while the progress of line-based features has been comparatively less notable. Among other works, we could highlight the approaches proposed by Br¨andli et al. [24], detecting and tracking line clusters in an event-by-event fashion, and Everding et al. [25], whose propose a plane ﬁtting approach on the event stream based on principal component analysis to track and cluster lines.
Processed_IDOL:_A_Framework_for_IMU-DVS_Odometry_using_Lines.txt	The proposed approach takes inspiration from previous continuous-time formulations and their integration with IMU measurements as in [18] and [19]. However, we also draw concepts from event-driven line tracking and clustering ap- proaches such as the methods described in [24] and [25], avoiding the need for supplementary sensing modalities as frame-based images and operating directly on the asyn- chronous event stream.
Processed_IDOL:_A_Framework_for_IMU-DVS_Odometry_using_Lines.txt	The asynchronism of event cameras represents a major technical challenge for state estimation. While originally fo- cused on discrete-time estimation, the use of rolling-shutter- like sensors (lidar, rolling-shutter camera, etc.), as well as multi-sensor platforms, lead the robotics community to de- velop continuous-time state estimation theory and methods. A large number of frameworks assume motion-models, often constant velocity, to interpolate the state variables in between discrete estimation timestamps [26], [27]. In [28], Furgale et al. introduce a fully continuous framework that considers the state as being the linear combination of temporal basis functions. Anderson and Barfoot [29] present a probabilistic approach to efﬁciently infer the state variables using Gaus- sian Process (GP) regression over a discrete maximum a posteriori estimation. A different paradigm is presented in [30], [31] where GPs are used as continuous representations of the inertial data allowing the characterization of the system’s pose in a continuous manner while relying on a discrete state. This is the approach employed in the proposed method with the use of Gaussian preintegrated measurements (GPMs) originally presented in [32].
Processed_IDOL:_A_Framework_for_IMU-DVS_Odometry_using_Lines.txt	camera, and both the system’s trajectory and the position of the 3D lines are simultaneously estimated. Addressing the asynchronicity of the event stream, a continuous repre- sentation of the inertial data is used to associate each event with inertial measurements. The state is then estimated by means of a discrete-state batch on-manifold optimization that accounts for the events individually.
Processed_Spectral_performance_of_SKA_Log-periodic_Antennas_.txt	ABSTRACT This paper is the ﬁrst in a series of papers describing the impact of antenna instru- mental artefacts on the 21-cm cosmology experiments to be carried out by the low frequency instrument (SKA1-LOW) of the Square Kilometre Array telescope (SKA), i.e., the Cosmic Dawn (CD) and the Epoch of Reionization (EoR). The smoothness of the passband response of the current log-periodic antenna being developed for the SKA1-LOW is analyzed using numerical electromagnetic simulations. The amplitude variations over the frequency range are characterized using low-order polynomials deﬁned locally, in order to study the impact of the passband smoothness in the instru- ment calibration and CD/EoR Science. A solution is oﬀered to correct a fast ripple found at 60 MHz during a test campaign at the SKA site at the Murchison Radio- astronomy Observatory, Western Australia in September 2015 with a minor impact on the telescope’s performance and design. A comparison with the Hydrogen Epoch of Reionization Array antenna is also shown demonstrating the potential use of the SKA1-LOW antenna for the Delay Spectrum technique to detect the EoR.
Processed_Spectral_performance_of_SKA_Log-periodic_Antennas_.txt	MHz) that will be used for the study of the early epochs of the Universe by looking at the redshifted signal from the 21- cm Hydrogen emission line. This instrument will be located in Western Australia. It will be a 512 station interferometer with baselines up to 60 km and more than 130,000 antenna elements (Turner 2015). Each station has been designed as a pseudo-random array of wide band antennas in order to min- imize the eﬀects of mutual coupling as shown by Gonzalez- Ovejero et al. (2011) and de Lera Acedo et al. (2011) and the eﬀects of side-lobes (Razavi-Ghods et al. 2012; El-makadema et al. 2014). Each station is 35 m in diameter in order to meet the desired beam-width (Mellema et al. 2013; Koop- mans et al. 2015).
Processed_Spectral_performance_of_SKA_Log-periodic_Antennas_.txt	Figure 1. Computer model of SKALA-2 antenna on top of a metallic ground plane (the ground plane is a mesh of wires with 30 cm pitch). The antenna is a log-periodic dipole array made of 4 identical metallic arms forming 2 polarizations with 9 dipoles each. The gray ring is the base of the antenna. The cabling coming down vertically in the center of the antenna is the optical ﬁber and power cable. The top enclosure with a green cap hosts the low-noise ampliﬁers and an RF-over-Fiber transmitter. The antenna has a footprint of 1.2 x 1.2 m and a height of 1.8 m.
Processed_Spectral_performance_of_SKA_Log-periodic_Antennas_.txt	de Lera Acedo et al. (2015b). SKALA-1 maximizes the tele- scope’s sensitivity over the ﬁeld of view across the 7:1 fre- quency band due to its moderate to high directivity and ﬂat impedance, in contrast to dipole antennas that have low dir- ectivity values and only show relatively ﬂat impedance over a narrow frequency band (typically 2:1 or 3:1 at most). Sub- sequently, a mechanical upgrade of SKALA-1 was designed and built, SKALA-2, in order to improve the design for mass manufacturing with negligible impact on the electromag- netic performance of the antenna. This upgraded design was presented in de Lera Acedo et al. (2015a). A computer model of the SKALA-2 antenna is shown in Figure 1. This paper describes the modiﬁcations to the length of the SKALA-2 antenna arms and component values of capacitors and in- ductors in the input matching network of the Low-Noise Ampliﬁer (LNA) in order to improve the spectral smooth- ness of the antenna system. The result of this improvement is SKALA-3.
Processed_Spectral_performance_of_SKA_Log-periodic_Antennas_.txt	This paper describes the analysis of the proposed SKA1- LOW antenna, SKALA-3, using a local low-order polyno- mial ﬁtting in order to assess its performance against real- istic CD/EoR science requirements. In Section 2 the require- ments derived in Trott & Wayth (2016) are summarized. Sec- tion 3 describes the current antenna topology and perform- ance. Section 4 focuses on the description of the analysis and the results from numerical simulations. Section 5 presents the improved bandpass smoothness to meet the science re- quirements of the proposed SKA1-LOW antenna. Section 6 shows a comparison with the Hydrogen Epoch of Reioniza- tion Array (HERA2) antenna against requirements derived for the Delay Spectrum technique for EoR observations. Fi- nally, Section 7 draws some conclusions.
Processed_Spectral_performance_of_SKA_Log-periodic_Antennas_.txt	Figure 2. Simpliﬁed RF schematic model of the ﬁrst ampliﬁcation stage of the LNA. The S-parameters and noise parameters for the transistors used in this design are obtained from the manufacturers. This diagram highlights the key components in the input side of the LNA that have been used for the improvement of the design (the DC blocking capacitors C1 and C2, the bias inductors L1, L2 and the ESD shunt inductors L3 and L4). The ESD diodes D1 and D2 are also shown here. For the simulations in this paper the full model of the LNA also included the second stage transistor although this has a small eﬀect on the matching with the antenna.
Processed_Spectral_performance_of_SKA_Log-periodic_Antennas_.txt	Table 1. Requirements from Trott & Wayth (2016). δ is the nor- malized fractional residual after a third-order polynomial ﬁtting.
Processed_Spectral_performance_of_SKA_Log-periodic_Antennas_.txt	requirements were also derived for second and fourth other polynomial ﬁttings). Strictly speaking this should be applied to the station beam passband. Our current analysis has been done at the antenna level, which will dominate the passband response of the instrument. Further analysis is needed to in- clude station beam eﬀects such as mutual coupling and other antenna eﬀects such as polarization and the instrument’s response across the whole ﬁeld of view (including spectral variations in the far ﬁeld pattern of the antenna).
Processed_Spectral_performance_of_SKA_Log-periodic_Antennas_.txt	In Trott & Wayth (2016) a formal procedure for the deriva- tion of spectral requirements for radio interferometers ded- icated to 21-cm cosmology experiments is presented. Here, we use the requirements derived in that paper to assess the design of the SKALA-2 and SKALA-3 antennas. In a com- panion paper (Trott et al. 2017, Paper II), we use realistic simulations to test the calibration performance of these an- tennas, and a comparison dipole (Tingay et al. 2013; Wayth et al. 2017, the Murchison Wideﬁeld Array - Engineering Development Array).
Processed_Spectral_performance_of_SKA_Log-periodic_Antennas_.txt	Furthermore, in Trott & Wayth (2016) requirements are derived for the necessary smoothness in the phase passband of the station’s voltage gain. Phase residuals aﬀect the ima- ging (tomography) experiment substantially. The so-called “phase noise” decorrelates the visibilities and adds an eﬀect- ive noise to the images that swamps the signal. So, in Trott & Wayth (2016) the maximum change in phase across any ﬁne channel (within the 100 - 200 MHz band) that would lead to swamping of the expected EoR signal is computed. In the worst case where the phase residual on each channel is ﬁxed over the full experiment, there cannot be more than a 0.04 degree change of phase over any ﬁne channel (4.58 kHz). An optimistic case where the phase residual is uncorrelated in time between calibration cycles, yielding an increase in dynamic range and weaker constraints, raises that limit to 0.2 degrees.
Processed_QB4AIRA:_A_Question_Bank_for_AI_Risk_Assessment.txt	The rapid advancement of Artificial Intelligence (AI), represented by ChatGPT, has raised concerns about responsible AI development and utilization. Existing frameworks lack a comprehensive synthesis of AI risk assessment questions. To address this, we introduce QB4AIRA, a novel question bank developed by refining questions from five globally recognized AI risk frameworks, categorized according to Australia’s AI ethics principles. QB4AIRA comprises 293 prioritized questions covering a wide range of AI risk areas, facilitating effective risk assessment. It serves as a valuable resource for stakeholders in assessing and managing AI risks, while paving the way for new risk frameworks and guidelines. By promoting responsible AI practices, QB4AIRA contributes to responsible AI deployment, mitigating potential risks and harms.
Processed_QB4AIRA:_A_Question_Bank_for_AI_Risk_Assessment.txt	Artificial Intelligence (AI) has witnessed remarkable advancements in recent years, permeating numerous domains and transforming various aspects of our lives. However, this rapid advancement of AI technologies has also raised concerns about their responsible development and application. For instance, while large language models, such as ChatGPT, have demonstrated impressive capabilities in generating text-based responses across a wide range of queries, they exhibit limitations and risks such as hallucination, generating harmful content, and user’s overreliance [1]. Incidents involving AI systems, documented in repositories like the AI Incident Database1, also highlight the urgency of addressing AI risks. As of July 2023, this repository has cataloged over 2,800 unfavorable incidents caused by AI systems worldwide, emphasizing the need for comprehensive risk management.
Processed_QB4AIRA:_A_Question_Bank_for_AI_Risk_Assessment.txt	To ensure the responsible development, deployment, and utilization of AI systems, the concept of Responsible AI (RAI) has emerged. RAI involves the practice of developing AI systems that benefit individuals, groups, and society as a whole, while minimizing the risk of adverse consequences [2, 3]. Risks in this context encompass the likelihood and consequences that arise when the development and use of AI systems fail to align with RAI principles, such as fairness, privacy, explainability, and accountability throughout the AI development process. Therefore, it is crucial to ensure that the development and use of AI systems adhere to these RAI principles.
Processed_QB4AIRA:_A_Question_Bank_for_AI_Risk_Assessment.txt	Playbook4 provides extensive guidance on identifying AI risks and recommending mitigations, but it lacks integration with diverse standards, frameworks, and related resources.
Processed_QB4AIRA:_A_Question_Bank_for_AI_Risk_Assessment.txt	To bridge this gap, we propose QB4AIRA as a comprehensive and connected resource for AI risk assessment. In this paper, we present the development of QB4AIRA by conducting a review of existing frameworks identified in [3], analyzing their core characteristics, and evaluating their alignment with RAI principles. We selected five frameworks to develop the question bank based on (i) global recognition, (ii) inclusion of risk assessment questions, and (iii) representation of different regions and industry leaders. These frameworks encompass various geopolitical contexts, including the EU Trustworthy AI Assessment List, Canada Algorithmic Impact Assessment5, Australia NSW AI Assurance Framework6, Microsoft (MS) Responsible AI Impact Assessment Guide7, and the US NIST AI Risk Management Framework.
Processed_QB4AIRA:_A_Question_Bank_for_AI_Risk_Assessment.txt	Through iterative refinement, we condensed 382 questions into 293, systematically categorized them based on Australia’s AI ethics principles [6], and conducted concept mapping to identify common risk themes [7]. The hierarchy of risk questions was structured using a decision tree [8], considering their level and sequence. QB4AIRA has undergone evaluation through two case studies.
Processed_QB4AIRA:_A_Question_Bank_for_AI_Risk_Assessment.txt	QB4AIRA is a collection of risk questions aimed at supporting the ethical development of AI systems. It is designed to meet the requirements of diverse users, including individuals at the C-level, project manager level, developer level, and beyond. QB4AIRA also tackles the distinct challenges and considerations linked to AI risk assessment and helps identify areas where risk management practices may be lacking.
Processed_QB4AIRA:_A_Question_Bank_for_AI_Risk_Assessment.txt	QB4AIRA is structured to ensure comprehensive coverage of key components and features a clear focus on RAI (Figure 1).
Processed_QB4AIRA:_A_Question_Bank_for_AI_Risk_Assessment.txt	Human, societal and environmental wellbeing, Human-centered Value, Fairness, Privacy and Security, Reliability and Safety, Transparency and Explainability, Contestability, and Accountability. To facilitate navigation and identification, a two-digit sequential identifier is assigned to each principle (P1 to P8).

Subsets and Splits

No community queries yet

The top public SQL queries from the community will appear here once available.