Dimitri Semenovich

Forecasting the onset of genocide and politicide: Annual out-of-sample forecasts on a global dataset, 1988–2003

We present what is, to the best of our knowledge, the first published set of annual out-of-sample forecasts of genocide and politicide based on a global dataset. Our goal is to produce a prototype for a real-time model capable of forecasting one year into the future. Building on the current literature, we take several important steps forward. We implement an unconditional two-stage model encompassing both instability and genocide, allowing our sample to be the available global data, rather than using conditional case selection or a case-control approach. We explore factors exhibiting considerable variance over time to improve yearly forecasting performance. And we produce annual lists of at-risk states in a format that should be of use to policymakers seeking to prevent such mass atrocities. Our out-of-sample forecasts for 1988–2003 predict 90.9% of genocide onsets correctly while also predicting 79.2% of non-onset years correctly, an improvement over a previous study using a case-control in-sample approach. We produce 16 annual forecasts based only on previous years’ data, which identify six of 11 cases of genocide/politicide onset within the top 5% of at-risk countries per year. We believe this represents substantial progress towards useful real-time forecasting of such rare events. We conclude by suggesting ways to further enhance predictive performance.

Dictionary learning framework for fabric defect detection

We present a new approach using dictionary learning framework to address textile fabric defect detection. Textile fabrics are textured materials whose images exhibit high periodicity among the repeated sub-patterns determined by weaving structure. Inspired by the image de-noising using the learned dictionary, we learn a dictionary from patches of textile fabric images, such a dictionary is able to approximate training samples well through a linear summation of its elements. Fabric defects can be regarded as a local anomaly against the relatively homogeneous texture. When modelling new samples with the dictionary learned from only the examples containing normal fabrics, the approximated version of an abnormal or defective sample will no longer contain defective region, resulting in a larger dissimilarity than a normal one, since the learned dictionary has been tuned to normal fabric structural features. Therefore, simply measuring the similarity between the original and its approximation is able to efficiently discriminate defective samples from normal, and a recently developed novelty detection algorithm, the support vector data description, is used to handle classification task. Experimental results show that the proposed algorithm can control both false alarm rate and missing detection rate within 5%, and extensions are also conducted.

Geometry aware local kernels for object recognition

Standard learning techniques can be difficult to apply in a setting where instances are sets of features, varying in cardinality and with additional geometric structure. Kernel-based classification methods can be effective in this situation as they avoid explicitly representing the instances. We describe a kernel function which attempts to establish correspondences between local features while also respecting the geometric structure. We generalize some of the existing work on context dependent kernels and demonstrate a connection to popular graph kernels. We also propose an efficient computation scheme which makes the new kernel applicable to instances with hundreds of features. The kernel function is shown to be positive semidefinite, making it suitable for use in a wide range of learning algorithms.

Sparse Dictionary Reconstruction for Textile Defect Detection

Inspired by the image de-noising techniques using learned dictionaries and sparse representation, we present a fabric defect detection scheme via sparse dictionary reconstruction. Fabric defects can be regarded as local anomalies against the relatively homogeneous texture background. Following from the flexibility of sparse representation, normal fabric samples can be efficiently represented using a linear combination of a few elements of a learned dictionary. When modeling new samples with a learned dictionary, tuned to the input data containing normal fabric structural features, abnormal or defective samples are likely to have larger dissimilarity than normal samples. We evaluate the proposed methods using ten different fabric types. Experimental results show that our method has many advantages in defect detection, especially in adapting variation of fabric textures.

Cell tracking and mitosis detection using splitting flow networks in phase-contrast imaging

Cell tracking is a crucial component of many biomedical image analysis applications. Many available cell tracking systems assume high precision of the cell detection module. Therefore low performance in cell detection can heavily affect the tracking results. Unfortunately cell segmentation modules often have significant errors, especially in the case of phase-contrast imaging. In this paper we propose a tracking method that does not rely on perfect cell segmentation and can deal with uncertainties by exploiting temporal information and aggregating the results of many frames. Our tracking algorithm is fully automated and can handle common challenges of tracking such as cells entering/exiting the screen and mitosis events. To handle the latter, we modify the standard flow network and introduce the concept of a splitting node into it. Experiment results show that adding temporal information from the video microscopy improves the cell/mitosis detection and results in a better tracking system.

Tensor Power Method for Efficient MAP Inference in Higher-order MRFs

We present a new efficient algorithm for maximizing energy functions with higher order potentials suitable for MAP inference in discrete MRFs. Initially we relax integer constraints on the problem and obtain potential label assignments using higher-order (tensor) power method. Then we utilise an ascent procedure similar to the classic ICM algorithm to converge to a solution meeting the original integer constraints.

Employing temporal information for cell segmentation using max-flow/min-cut in phase-contrast video microscopy

Cell segmentation is a crucial step in many bio-medical image analysis applications and it can be considered as an important part of a tracking system. Segmentation in phase-contrast images is a challenging task since in this imaging technique, the background intensity is approximately similar to the cell pixel intensity. In this paper we propose an interactive automatic pixel level segmentation algorithm, that uses temporal information to improve the segmentation result. This algorithm is based on the max-flow/min-cut algorithm and can be solved in polynomial time. This method is not restricted to any specific cell shape and segments cells of various shapes and sizes. The results of the proposed algorithm show that using the temporal information does improve segmentation considerably.

A spectral method for context based disambiguation of image annotations

In this work we employ contextual information to improve the quality of image labellings provided by an existing automatic image annotation algorithm in a weakly supervised setting, where each training image is labelled but it is not known which part of the image its labels are referring to. We recast the problem into that of constructing a graph which encodes pairwise consistency of candidate annotations and observe that mutually consistent labels will form a compact cluster in this graph. We recover the clusters using a spectral theory based technique. The results are demonstrated on the Corel5k dataset. With improvements in the range of 25%–55% the performance in some cases approaches the state of the art despite using a very simple base algorithm.

Elections and Political Instability: Ballots to Bullets, Voting to Violence?

Serious instability and deadly political violence have surrounded several recent elections and gained global media coverage. Do elections in general make such incidents more likely? Are elections especially dangerous in partially democratic or ethnically divided states? Elections are among the most common and well-studied of political events. Nevertheless the academic literature is divided on this topic. Our quantitative analysis suggests that elections are rarely dangerous, even when they occur in difficult political or ethnic contexts. On the contrary, we find that elections tend to decrease the chance of instability in states with high ethnic fractionalization. We argue that during election periods, high levels of ethnic fractionalization lower the expected payoffs for violent methods of political gain relative to the expected value of minority, opposition or coalition status. Elections, we posit, provide inducements that are particularly appealing to ethnic groups seeking limited power or autonomy within a state, while such limited gains through institutionalized mechanisms controlled by the state are also less threatening to incumbent governments or other ethnic groups in the society. We point to the importance of both our statistically significant and insignificant findings. We conclude by drawing out some theoretical and policy implications of the lack of evidence for a general effect for elections, and the apparently robust evidence for a pacifying effect in ethnically divided states.

Political Competition and the Initiation of International Conflict: A New Perspective on the Institutional Foundations of Democratic Peace

Although some scholars claim that the empirical evidence for the very low instance of interstate war between democracies is well established, others have raised new challenges. But even if democratic peace is observed, its theoretical explanation remains unresolved. Consensus has not emerged among competing approaches, some of which are criticized for offering monadic logic for a dyadic phenomenon. This article synthesizes recent literature to advance a simple, but distinct, explicitly dyadic theory about institutionalized political competition, leading to expectations that it is the most important source of democratic peace. While the authors are far from the first to consider political competition, their approach stands out in according it the central role in a dyadic theory focused on the regime type of initiators and target states. They argue that potential vulnerability to opposition criticism on target-regime-specific normative and costs-of-war bases is more fundamental than mechanisms such as audience costs, informational effects, or public goods logic. Incumbents in high-competition states will be reluctant to initiate conflict with a democracy due to anticipated inability to defend the conflict as right, necessary, and winnable. The authors present new and highly robust evidence that democratic peace is neither spurious nor a methodological artifact, and that it can be attributed to high-competition states’ aversion to initiating fights with democracies.