Sparse Datasets Research Articles

TT-SVD: An Efficient Sparse Decision-Making Model With Two-Way Trust Recommendation in the AI-Enabled IoT Systems

The convergence of AI and IoT enables data to be quickly explored and turned into vital decisions, and however, there are still some challenging issues to be further addressed. For example, lacking of enough data in AI-based decision making [so-called sparse decision making (SDM)] will decrease the efficiency dramatically, or even disable the intelligent IoT networks. Taking the intelligent IoT networks as the network infrastructure, the recommendation systems have been facing such SDM problems. A naive solution is to introduce trust information. However, trust information may also face the difficulty of sparse trust evidence (also known as sparse trust problem). In our work, an accurate SDM model with two-way trust recommendation in the AI-enabled IoT systems is proposed, named TT-SVD. Our model incorporates both trust information and rating information more thoroughly, which can efficiently alleviate the above-mentioned sparse trust problem and therefore be able to solve the cold start and data sparsity problems. Specifically, we first consider the twofold trust influences from both trustees and trusters, which can be represented by a factor named trust propensity. To this end, we propose a dual model, including a truster model (TrusterSVD) and a trustee model (TrusteeSVD) based on an existing rating-only recommendation model called SVD++, which are integrated by the weighted average and yield the final model, TT-SVD. The experimental results show that our model outperforms the state-of-the-art, including SVD and TrustSVD in both the “all users” and “cold start users” cases, and the accuracy improvement can reach a maximum of 29%. Complexity analysis shows that our model is equally suitable for the case of large sparse data sets. In summary, our model can effectively solve the sparse decision problem by introducing the two-way trust recommendation, and hence improve the efficiency of the intelligent recommendation systems.

Retrieval of rapeseed leaf area index using the PROSAIL model with canopy coverage derived from UAV images as a correction parameter

Leaf area index (LAI), which is an important structural parameter, plays a vital role in evaluating crop growth and yield. In this study, we used the canopy coverage (CC) derived from unmanned aerial vehicle (UAV) images as a correction parameter in the PROSAIL model coupled with a neural network (NN) to improve the accuracy of LAI inversion of rapeseed plots. CC had a significantly positive impact on the accuracy of LAI inversion especially in sparse canopy structure with the 22.24% decrease in the entire dataset and 35.76% decrease in the sparse canopy dataset. We then compared the inversion performances of an empirical statistical model (ESM) based on a vegetation index and the PROSAIL model incorporating CC correction for 2016 and 2018 datasets. The ESM performed better in modeling the 2016 dataset, but its accuracy was much lower for the 2018 dataset (2016: NRMSE = 0.131; 2018: NRSME = 0.348). Overall, the PROSAIL model was more robust over these two datasets (2016: NRMSE = 0.152; 2018: NRMSE = 0.168). In addition, the original-resolution images were resampled to six coarse resolutions to evaluate the influence of image resolution on the LAI inversion performance of the PROSAIL model. When pixel size increased to more than 10 cm, the inversion accuracy began to decrease dramatically. In conclusion, introducing a canopy coverage correction parameter in the PROSAIL model improved its performance in retrieving rapeseed LAI.

Revisiting Co-Occurring Directions: Sharper Analysis and Efficient Algorithm for Sparse Matrices

We study the streaming model for approximate matrix multiplication (AMM). We are interested in the scenario that the algorithm can only take one pass over the data with limited memory. The state-of-the-art deterministic sketching algorithm for streaming AMM is the co-occurring directions (COD), which has much smaller approximation errors than randomized algorithms and outperforms other deterministic sketching methods empirically. In this paper, we provide a tighter error bound for COD whose leading term considers the potential approximate low-rank structure and the correlation of input matrices. We prove COD is space optimal with respect to our improved error bound. We also propose a variant of COD for sparse matrices with theoretical guarantees. The experiments on real-world sparse datasets show that the proposed algorithm is more efficient than baseline methods.

Valve Detection for Autonomous Water Pipeline Inspection Platform

Water distribution and transmission lines are indispensable to urban infrastructure. The water pipelines are subject to both structural and functional deterioration due to various reasons including aging, negligence, and high demand for water supply. Hence, to ensure a safe and reliable water supply, the water utilities need to perform routine pipe condition assessment. The condition assessment is usually carried out by visual inspection with the machine vision system carried by a robotic platform. The inspection platforms will capture the internal condition of the water pipelines in a video stream. However, the robotic platform frequently experiences difficulties while traversing through the valves installed along the pipeline. This inhibits and disrupts the inspection process of the water pipelines. Therefore, this article proposes a deep learning-based automatic valve detection framework to facilitate the robot’s navigation and ensure continuous inspection without any interruptions. The valve detection model is developed by combining MobileNet-160 and Feature Pyramid Network and is named as MFPN. The developed framework also employs a generative adversarial network to solve the sparse dataset issues and improve the generality of the framework. The comparative study and ablation analyses demonstrate that the proposed framework can achieve a higher mAP value of 89.11% in comparison with the state of the art. Hence, this light weight and efficient solution can be deployed to the robotic platform for real-time valve detection and enable autonomous navigation of the robotic platform for condition assessment of water pipelines.

S-MPEC: Sparse Matrix Multiplication Performance Estimator on a Cloud Environment

Sparse matrix multiplication (SPMM) is widely used for various machine learning algorithms. As the applications of SPMM using large-scale datasets become prevalent, executing SPMM jobs on an optimized setup has become very important. Execution environments of distributed SPMM tasks on cloud resources can be set up in diverse ways with respect to the input sparse datasets, distinct SPMM implementation methods, and the choice of cloud instance types. In this paper, we propose S-MPEC which can predict latency to complete various SPMM tasks using Apache Spark on distributed cloud environments. We first characterize various distributed SPMM implementations on Apache Spark. Considering the characters and hardware specifications on the cloud, we propose unique features to build a GB-regressor model and Bayesian optimizations. Our proposed S-MPEC model can predict latency on an arbitrary SPMM task accurately and recommend an optimal implementation method. Thorough evaluation of the proposed system reveals that a user can expect 44% less latency to complete SPMM tasks compared with the native SPMM implementations in Apache Spark.

Comparison of 2-D and 3-D full waveform inversion imaging using wide-angle seismic data from the Deep Galicia Margin

SUMMARY Full waveform inversion (FWI) is a data-fitting technique capable of generating high-resolution velocity models with a resolution down to half the seismic wavelength. FWI is applied typically to densely sampled seismic data. In this study, we applied FWI to 3-D wide-angle seismic data acquired using sparsely spaced ocean bottom seismometers (OBSs) from the Deep Galicia Margin west of Iberia. Our data set samples the S-reflector, a low-angle detachment present in this area. Here we highlight differences between 2-D, 2.5-D and 3-D-FWI performances using a real sparsely spaced data set. We performed 3-D FWI in the time domain and compared the results with 2-D and 2.5-D FWI results from a profile through the 3-D model. When overlaid on multichannel seismic images, the 3-D FWI results constrain better the complex faulting within the pre- and syn-rift sediments and crystalline crust compared to the 2-D result. Furthermore, we estimate variable serpentinization of the upper mantle below the S-reflector along the profile using 3-D FWI, reaching a maximum of 45 per cent. Differences in the data residuals of the 2-D, 2.5-D and 3-D inversions suggest that 2-D inversion can be prone to overfitting when using a sparse data set. To validate our results, we performed tests to recover the anomalies introduced by the inversions in the final models using synthetic data sets. Based on our comparison of the velocity models, we conclude that the use of 3-D data can partially mitigate the problem of receiver sparsity in FWI.

Gravity Control-Based Data Augmentation Technique for Improving VR User Activity Recognition

The neural-network-based human activity recognition (HAR) technique is being increasingly used for activity recognition in virtual reality (VR) users. The major issue of a such technique is the collection large-scale training datasets which are key for deriving a robust recognition model. However, collecting large-scale data is a costly and time-consuming process. Furthermore, increasing the number of activities to be classified will require a much larger number of training datasets. Since training the model with a sparse dataset can only provide limited features to recognition models, it can cause problems such as overfitting and suboptimal results. In this paper, we present a data augmentation technique named gravity control-based augmentation (GCDA) to alleviate the sparse data problem by generating new training data based on the existing data. The benefits of the symmetrical structure of the data are that it increased the number of data while preserving the properties of the data. The core concept of GCDA is two-fold: (1) decomposing the acceleration data obtained from the inertial measurement unit (IMU) into zero-gravity acceleration and gravitational acceleration, and augmenting them separately, and (2) exploiting gravity as a directional feature and controlling it to augment training datasets. Through the comparative evaluations, we validated that the application of GCDA to training datasets showed a larger improvement in classification accuracy (96.39%) compared to the typical data augmentation methods (92.29%) applied and those that did not apply the augmentation method (85.21%).

SHAPE Directed Discovery of New Functions in Large RNAs.

RNA lies upstream of nearly all biology and functions as the central conduit of information exchange in all cells. RNA molecules encode information both in their primary sequences and in complex structures that form when an RNA folds back on itself. From the time of discovery of mRNA in the late 1950s until quite recently, we had only a rudimentary understanding of RNA structure across vast regions of most messenger and noncoding RNAs. This deficit is now rapidly being addressed, especially by selective 2'-hydroxyl acylation analyzed by primer extension (SHAPE) chemistry, mutational profiling (MaP), and closely related platform technologies that, collectively, create chemical microscopes for RNA. These technologies make it possible to interrogate RNA structure, quantitatively, at nucleotide resolution, and at large scales, for entire mRNAs, noncoding RNAs, and viral RNA genomes. By applying comprehensive structure probing to diverse problems, we and others are showing that control of biological function mediated by RNA structure is ubiquitous across prokaryotic and eukaryotic organisms.Work over the past decade using SHAPE-based analyses has clarified key principles. First, the method of RNA structure probing matters. SHAPE-MaP, with its direct and one-step readout that probes nearly every nucleotide by reaction at the 2'-hydroxyl, gives a more detailed and accurate readout than alternatives. Second, comprehensive chemical probing is essential. Focusing on fragments of large RNAs or using meta-gene or statistical analyses to compensate for sparse data sets misses critical features and often yields structure models with poor predictive power. Finally, every RNA has its own internal structural personality. There are myriad ways in which RNA structure modulates sequence accessibility, protein binding, translation, splice-site choice, phase separation, and other fundamental biological processes. In essentially every instance where we have applied rigorous and quantitative SHAPE technologies to study RNA structure-function interrelationships, new insights regarding biological regulatory mechanisms have emerged. RNA elements with more complex higher-order structures appear more likely to contain high-information-content clefts and pockets that bind small molecules, broadly informing a vigorous field of RNA-targeted drug discovery.The broad implications of this collective work are twofold. First, it is long past time to abandon depiction of large RNAs as simple noodle-like or gently flowing molecules. Instead, we need to emphasize that nearly all RNAs are punctuated with distinctive internal structures, a subset of which modulate function in profound ways. Second, structure probing should be an integral component of any effort that seeks to understand the functional nexuses and biological roles of large RNAs.

On Addressing the Low Rating Prediction Coverage in Sparse Datasets Using Virtual Ratings

On Addressing the Low Rating Prediction Coverage in Sparse Datasets Using Virtual Ratings

Materials Fingerprinting Classification

Significant progress in many classes of materials could be made with the availability of experimentally-derived large datasets composed of atomic identities and three-dimensional coordinates. Methods for visualizing the local atomic structure, such as atom probe tomography (APT), which routinely generate datasets comprised of millions of atoms, are an important step in realizing this goal. However, state-of-the-art APT instruments generate noisy and sparse datasets that provide information about elemental type, but obscure atomic structures, thus limiting their subsequent value for materials discovery. The application of a materials fingerprinting process, a machine learning algorithm coupled with topological data analysis, provides an avenue by which here-to-fore unprecedented structural information can be extracted from an APT dataset. As a proof of concept, the material fingerprint is applied to high-entropy alloy APT datasets containing body-centered cubic (BCC) and face-centered cubic (FCC) crystal structures. A local atomic configuration centered on an arbitrary atom is assigned a topological descriptor, with which it can be characterized as a BCC or FCC lattice with near perfect accuracy, despite the inherent noise in the dataset. This successful identification of a fingerprint is a crucial first step in the development of algorithms which can extract more nuanced information, such as chemical ordering, from existing datasets of complex materials. Program summaryProgram Title: Materials FingerprintingCPC Library link to program files:https://doi.org/10.17632/2fhch3x85m.1Developer's repository link:https://github.com/maroulaslab/Materials-FingerprintingLicensing provisions: GPLv3Programming language: PythonSupplementary material: A user manual and examples are provided with the source code in the GitHub repository.Nature of problem: Atom probe tomography provides sub-nanometer resolution of a material, but due to noise and sparsity introduced by the process, the crystal structure of a material cannot presently be determined from the resulting data.Solution method: Our Materials Fingerprinting library presents a topologically informed machine learning methodology to classify the lattice structure of a material from atomic probe tomography data. We create persistence diagrams from small neighborhoods centered at each atom in the resulting APT data and use the summary statistics of a novel metric on the space of persistence diagrams as features for a classification algorithm.

Weighted multi-view co-clustering (WMVCC) for sparse data

Multi-view clustering has gained importance in recent times due to the large-scale generation of data, often from multiple sources. Multi-view clustering refers to clustering a set of objects which are expressed by multiple set of features, known as views, such as movies being expressed by the list of actors or by a textual summary of its plot. Co-clustering, on the other hand, refers to the simultaneous grouping of data samples and features under the assumption that samples exhibit a pattern only under a subset of features. This paper combines multi-view clustering with co-clustering and proposes a new Weighted Multi-View Co-Clustering (WMVCC) algorithm. The motivation behind the approach is to use the diversity of features provided by multiple sources of information while exploiting the power of co-clustering. The proposed method expands the clustering objective function to a unified co-clustering objective function across all the multiple views. The algorithm follows the k-means strategy and iteratively optimizes the clustering by updating cluster labels, features, and view weights. A local search is also employed to optimize the clustering result using weighted multi-step paths in a graph. Experiments are conducted on several benchmark datasets. The results show that the proposed approach converges quickly, and the clustering performance significantly outperforms other recent and state-of-the-art algorithms on sparse datasets.

LoBSTr: Real‐time Lower‐body Pose Prediction from Sparse Upper‐body Tracking Signals

AbstractWith the popularization of games and VR/AR devices, there is a growing need for capturing human motion with a sparse set of tracking data. In this paper, we introduce a deep neural network (DNN) based method for real‐time prediction of the lower‐body pose only from the tracking signals of the upper‐body joints. Specifically, our Gated Recurrent Unit (GRU)‐based recurrent architecture predicts the lower‐body pose and feet contact states from a past sequence of tracking signals of the head, hands, and pelvis. A major feature of our method is that the input signal is represented by the velocity of tracking signals. We show that the velocity representation better models the correlation between the upper‐body and lower‐body motions and increases the robustness against the diverse scales and proportions of the user body than position‐orientation representations. In addition, to remove foot‐skating and floating artifacts, our network predicts feet contact state, which is used to post‐process the lower‐body pose with inverse kinematics to preserve the contact. Our network is lightweight so as to run in real‐time applications. We show the effectiveness of our method through several quantitative evaluations against other architectures and input representations with respect to wild tracking data obtained from commercial VR devices.

Zero-Inflated Poisson Distribution of Sedimented Cells in Multi-Layered Microwell Arrays

Open arrays of micro-scale wells (microwells; MW) are a popular platform for trapping biological cells, as they are gentler than other methods and their openness circumvents several problems associated with enclosed alternatives. This paper presents a dual-layered polymeric film featuring an imprinted MW array (MWA) and various complimentary shallower features that streamline both optical microscopy and alignment with an immunobiosensing (IBS) slide. The dual-layered MWA design presented in this paper represents a substantial improvement over our previous designs. The most substantial contribution of this paper lies with its statistical analysis of the trapped cell count datasets obtained from experiments using this refined MWA design. This analysis confirms experimentally that the distribution of cells into a MWA following sedimentation is indeed naturally Poisson distributed. Moreover, this analysis also shows that a zero-inflated Poisson (ZIP) distribution provides a superior fit, by incorporating an additional variable quantifying dataset sparsity. Furthermore, it is shown that maximum likelihood estimators (MLEs) for the parameters of these Poisson fits are superior to method of moments-based alternatives. This paper should prove useful for those seeking to develop a MWA with which to trap cells via sedimentation, and to mathematically describe this trapping process.

Random vector functional link neural network based ensemble deep learning

In this paper, we propose deep learning frameworks based on the randomized neural network. Inspired by the principles of Random Vector Functional Link (RVFL) network, we present a deep RVFL network (dRVFL) with stacked layers. The parameters of the hidden layers of the dRVFL are randomly generated within a suitable range and kept fixed while the output weights are computed using the closed-form solution as in a standard RVFL network. We also propose an ensemble deep network (edRVFL) that can be regarded as a marriage of ensemble learning with deep learning. Unlike traditional ensembling approaches that require training several models independently from scratch, edRVFL is obtained by training a single dRVFL network once. Both dRVFL and edRVFL frameworks are generic and can be used with any RVFL variant. To illustrate this, we integrate the deep learning RVFL networks with a recently proposed sparse pre-trained RVFL (SP-RVFL). Experiments on 46 tabular UCI classification datasets and 12 sparse datasets demonstrate that the proposed deep RVFL networks outperform state-of-the-art deep feed-forward neural networks (FNNs).

Ensemble smoother with multiple data assimilation to simultaneously estimate the source location and the release history of a contaminant spill in an aquifer

The source location and the time history of a pollutant released in an aquifer are very relevant information for the design of effective remediation strategies. Usually, their identification requires solving an inverse problem when the only available information about the groundwater contamination event is a sparse set of concentration data collected in the aquifer at a few points downstream from the source. Here, a novel approach is proposed to solve the inverse problem: the use of the Ensemble Smoother with Multiple Data Assimilation (ES-MDA) in the context of source contamination identification. This method is used for the simultaneous determination of the time history and the source location of a pollutant release based on observed concentration data and a calibrated numerical model of groundwater flow and mass transport in the aquifer. The ES-MDA is demonstrated in two case studies. The first one is based on an analytical solution of the flow and transport equations, aimed at the estimation of the source location and the release history of a nonreactive pollutant spreading in a two-dimensional homogeneous aquifer from a point source. For this case, different alternatives are considered for the spatial distribution of the observation points, the concentration sampling frequency, the ensemble size and the use of covariance localization and covariance inflation techniques in the formulation of the smoother. The purpose of this case is to test the new approach, analyze its performance and also to identify the conditions that render the problem ill-posed and, therefore, without solution; also, in this case, a new spatiotemporal iterative localization is presented. In the second case study, we use real data collected in a laboratory sandbox that reproduces a vertical cross-section of an unconfined aquifer with two-dimensional quasi-parallel flow between constant-head boundaries. The results show that the location, time and number of observations, the ensemble size and the application of covariance localization and covariance inflation techniques have an impact on the final solution. A well-designed monitoring network and the application of covariance corrections improve the performance of the ES-MDA and help avoiding ill-posedness and equifinality. The application to laboratory data validates the potential of ES-MDA to simultaneously estimate the time history and the source location of a pollutant released in groundwater in real cases.

Intelligent wind farm control via deep reinforcement learning and high-fidelity simulations

Wind farms’ power-generation efficiency is constrained by the high system complexity. A novel deep reinforcement learning (RL)-based wind farm control scheme is proposed to handle this challenge and achieve power generation optimization. A reward regularization (RR) module is designed to estimate wind turbines’ normalized power outputs under different yaw settings and uncertain wind conditions, which brings strong robustness and adaptability to the proposed control scheme. The RR module is then combined with the deep deterministic policy gradient algorithm to evaluate the optimal yaw settings for all the wind turbines within the farm. The proposed wind farm control scheme is data-driven and model-free, which addresses the limitations of current approaches, including reliance on accurate analytical/parametric models and lack of adaptability to uncertain wind conditions. In addition, a novel composite learning-based controller for each turbine is designed to achieve closed-loop yaw tracking, which can guarantee the exponential convergence of tracking errors in the presence of uncertainties of yaw actuators. The whole control system can be pre-trained offline and fine-tuned online, providing an easy-to-apply solution with enhanced generality and flexibility for wind farms. High-fidelity simulations with SOWFA (simulator for offshore wind farm applications) and Tensorflow show that the proposed scheme can significantly improve the wind farm’s power generation by exploiting a sparse data set without requiring any wake model.

Efficient algorithms for deriving complete frequent itemsets from frequent closed itemsets

When mining frequent itemsets (abbr. FIs) from dense datasets, it usually produces too many itemsets and results in the mining task to suffer from a very long execution time and high memory consumption. Frequent closed itemset (abbr. FCI) is a compact and lossless representation of FI. Mining FCIs can not only reduce the execution time and memory usage, but also reserve the complete information of FIs derived from FCIs. Although many studies have been proposed with various efficient methods for mining FCIs, few of them have developed algorithms for efficiently deriving FIs from FCIs. In this work, we propose two efficient algorithms named DFI-List and DFI-Growth for efficiently deriving FIs from FCIs. The both algorithms adopt depth-first search and divide-and-conquer methodology to derive all the FIs. DFI-List efficiently derives all the FIs with a vertical index structure called Cid List. DFI-Growth compresses the information of FCIs into tree structures and applies pattern-growth strategy to derive FIs from the trees. Empirical experiments show that DFI-List is the most efficient and scalable algorithm on the dense datasets. For example, when the minimum support threshold is set to 50% on the Chess dataset, DFI-List runs faster than LevelWise (Pasquier et al. Inf Syst 24(1): 25-46, 1999b) over 100 times. As for DFI-Growth, it is the most stable and memory efficient algorithm on the sparse datasets. Both DFI-Growth and DFI-List are superior to the state-of-the-art algorithm (Pasquier et al. Inf Syst 24(1): 25-46, 199b) in terms of execution time.

Regularized bi-directional co-clustering

The simultaneous clustering of documents and words, known as co-clustering, has proved to be more effective than one-sided clustering in dealing with sparse high-dimensional datasets. By their nature, text data are also generally unbalanced and directional. Recently, the von Mises–Fisher (vMF) mixture model was proposed to handle unbalanced data while harnessing the directional nature of text. In this paper, we propose a general co-clustering framework based on a matrix formulation of vMF model-based co-clustering. This formulation leads to a flexible framework for text co-clustering that can easily incorporate both word–word semantic relationships and document–document similarities. By contrast with existing methods, which generally use an additive incorporation of similarities, we propose a bi-directional multiplicative regularization that better encapsulates the underlying text data structure. Extensive evaluations on various real-world text datasets demonstrate the superior performance of our proposed approach over baseline and competitive methods, both in terms of clustering results and co-cluster topic coherence.

WEDGE: imputation of gene expression values from single-cell RNA-seq datasets using biased matrix decomposition.

The low capture rate of expressed RNAs from single-cell sequencing technology is one of the major obstacles to downstream functional genomics analyses. Recently, a number of imputation methods have emerged for single-cell transcriptome data, however, recovering missing values in very sparse expression matrices remains a substantial challenge. Here, we propose a new algorithm, WEDGE (WEighted Decomposition of Gene Expression), to impute gene expression matrices by using a biased low-rank matrix decomposition method. WEDGE successfully recovered expression matrices, reproduced the cell-wise and gene-wise correlations and improved the clustering of cells, performing impressively for applications with sparse datasets. Overall, this study shows a potent approach for imputing sparse expression matrix data, and our WEDGE algorithm should help many researchers to more profitably explore the biological meanings embedded in their single-cell RNA sequencing datasets. The source code of WEDGE has been released at https://github.com/QuKunLab/WEDGE.

Spatiotemporal variability in lithic technology of Middle-to-Upper Paleolithic Asia: A new dataset and its statistical analyses

The dispersal processes of anatomically modern humans in Asia during the late Middle to early Upper Paleolithic periods have been a focus of intensive research in recent years through interdisciplinary approaches including genetic, paleoanthropological, and archeological studies. However, sparse datasets, often only available from limited time-space contexts have not produced a satisfactory picture of the vast geographic extension of this continent. This study investigated how the chrono-spatial patterning of lithic technological records, which are widely available across Asia, might contribute to our better understanding of the population dynamics in this period. The study is based on a large database (PaleoAsiaDB) recently developed by the authors to accommodate the relevant datasets, including geographical coordinates of each site and available radiometric dates, for the Middle and Upper Paleolithic periods in Asia. Each of the 895 lithic assemblages was characterized in terms of the presence/absence of 24 lithic modes/submodes that were adapted from the work of John Shea to fit our research objective. The dataset was analyzed using exploratory statistical methods, including principal component analysis. The results revealed marked spatiotemporal patterns in technological variability from the late Middle to early Upper Paleolithic periods. Some can be interpreted as a result of modern human dispersal, whereas others are better explained from ecological and other perspectives rather than authorships. Although this study has obvious limitations due to its reliance on the information sourced from the literature, it provides a global synthesis of lithic variability in the crucial period of the dispersals of modern humans on a continental scale for the first time. The results serve as a guide to be tested with evidence from other research strategies, such as genetics and fossil studies.