Modeling Efficient Classification as a Process of Confidence Assessment and Delegation

In visual object detection and recognition, classifiers have two interesting characteristics: accuracy and speed. Accuracy depends on the complexity of the image features and classifier decision surfaces. Speed depends on the hardware and the computational effort required to use the features and decision surfaces. When attempts to increase accuracy lead to increases in complexity and effort, it is necessary to ask how much are we willing to pay for increased accuracy. For example, if increased computational effort implies quickly diminishing returns in accuracy, then those designing inexpensive surveillance applications cannot aim for maximum accuracy at any cost. It becomes necessary to find trade-offs between accuracy and effort. We study efficient classification of images depicting real-world objects and scenes. Classification is efficient when a classifier can be controlled so that the desired trade-off between accuracy and effort (speed) is achieved and unnecessary computations are avoided on a per input basis. A framework is proposed for understanding and modeling efficient classification of images. Classification is modeled as a tree-like process. In designing the framework, it is important to recognize what is essential and to avoid structures that are narrow in applicability. Earlier frameworks are lacking in this regard. The overall contribution is two-fold. First, the framework is presented, subjected to experiments, and shown to be satisfactory. Second, certain unconventional approaches are experimented with. This allows the separation of the essential from the conventional. To determine if the framework is satisfactory, three categories of questions are identified: trade-off optimization, classifier tree organization, and rules for delegation and confidence modeling. Questions and problems related to each category are addressed and empirical results are presented. For example, related to trade-off optimization, we address the problem of computational bottlenecks that limit the range of trade-offs. We also ask if accuracy versus effort trade-offs can be controlled after training. For another example, regarding classifier tree organization, we first consider the task of organizing a tree in a problem-specific manner. We then ask if problem-specific organization is necessary.

Structure and Dynamics of Coil-like Molecules by Residual Dipolar Couplings

Nuclear magnetic resonance (NMR) spectroscopy provides us with many means to study biological macromolecules in solution. Proteins in particular are the most intriguing targets for NMR studies. Protein functions are usually ascribed to specific three-dimensional structures but more recently tails, long loops and non-structural polypeptides have also been shown to be biologically active. Examples include prions, -synuclein, amylin and the NEF HIV-protein. However, conformational preferences in coil-like molecules are difficult to study by traditional methods. Residual dipolar couplings (RDCs) have opened up new opportunities; however their analysis is not trivial. Here we show how to interpret RDCs from these weakly structured molecules. The most notable residual dipolar couplings arise from steric obstruction effects. In dilute liquid crystalline media as well as in anisotropic gels polypeptides encounter nematogens. The shape of a polypeptide conformation limits the encounter with the nematogen. The most elongated conformations may come closest whereas the most compact remain furthest away. As a result there is slightly more room in the solution for the extended than for the compact conformations. This conformation-dependent concentration effect leads to a bias in the measured data. The measured values are not arithmetic averages but essentially weighted averages over conformations. The overall effect can be calculated for random flight chains and simulated for more realistic molecular models. Earlier there was an implicit thought that weakly structured or non-structural molecules would not yield to any observable residual dipolar couplings. However, in the pioneering study by Shortle and Ackerman RDCs were clearly observed. We repeated the study for urea-denatured protein at high temperature and also observed indisputably RDCs. This was very convincing to us but we could not possibly accept the proposed reason for the non-zero RDCs, namely that there would be some residual structure left in the protein that to our understanding was fully denatured. We proceeded to gain understanding via simulations and elementary experiments. In measurements we used simple homopolymers with only two labelled residues and we simulated the data to learn more about the origin of RDCs. We realized that RDCs depend on the position of the residue as well as on the length of the polypeptide. Investigations resulted in a theoretical model for RDCs from coil-like molecules. Later we extended the studies by molecular dynamics. Somewhat surprisingly the effects are small for non-structured molecules whereas the bias may be large for a small compact protein. All in all the work gave clear and unambiguous results on how to interpret RDCs as structural and dynamic parameters of weakly structured proteins.

Ischemic Stroke in Young Adults

Stroke is the second leading cause of death and the leading cause of disability worldwide. Of all strokes, up to 80% to 85% are ischemic, and of these, less than 10% occur in young individuals. Stroke in young adults—most often defined as stroke occurring under the age of 45 or 50—can be particularly devastating due to long expected life-span ahead and marked socio-economic consequences. Current basic knowledge on ischemic stroke in this age group originates mostly from rather small and imprecise patient series. Regarding emergency treatment, systematic data on use of intravenous thrombolysis are absent. For this Thesis project, we collected detailed clinical and radiological data on all consecutive patients aged 15 to 49 with first-ever ischemic stroke between 1994 and 2007 treated at the Helsinki University Central Hospital. The aims of the study were to define demographic characteristics, risk factors, imaging features, etiology, and long-term mortality and its predictors in this patient population. We additionally sought to investigate, whether intravenous thrombolysis is safe and beneficial for the treatment of acute ischemic stroke in the young. Of our 1008 patients, most were males (ratio 1.7:1), who clearly outnumbered females after the age of 44, but females were preponderant among those aged <30. Occurrence increased exponentially. The most frequent risk factors were dyslipidemia (60%), smoking (44%), and hypertension (39%). Risk factors accumulated in males and along aging. Cardioembolism (20%) and cervicocerebral artery dissection (15%) were the most frequent etiologic subgroups, followed by small-vessel disease (14%), and large-artery atherosclerosis (8%). A total of 33% had undetermined etiology. Left hemisphere strokes were more common in general. Posterior circulation infarcts were more common among those aged <45. Multiple brain infarcts were present in 23% of our patients, 13% had silent infarcts, and 5% had leukoaraiosis. Of those with silent brain infarcts, majority (54%) had only a single lesion, and most of the silent strokes were located in basal ganglia (39%) and subcortical regions (21%). In a logistic regression analysis, type 1 diabetes mellitus in particular predicted the presence of both silent brain infarcts (odds ratio 5.78, 95% confidence interval 2.37-14.10) and leukoaraiosis (9.75; 3.39-28.04). We identified 48 young patients with hemispheric ischemic stroke treated with intravenous tissue plasminogen activator, alteplase. For comparisons, we searched 96 untreated control patients matched by age, gender, and admission stroke severity, as well as 96 alteplase-treated older controls aged 50 to 79 matched by gender and stroke severity. Alteplase-treated young patients recovered more often completely (27% versus 10%, P=0.010) or had only mild residual symptoms (40% versus 22%, P=0.025) compared to age-matched controls. None of the alteplase-treated young patients had symptomatic intracerebral hemorrhage or died within 3-month follow-up. Overall long-term mortality was low in our patient population. Cumulative mortality risks were 2.7% (95% confidence interval 1.5-3.9%) at 1 month, 4.7% (3.1-6.3%) at 1 year, and 10.7% (9.9-11.5%) at 5 years. Among the 30-day survivors who died during the 5-year follow-up, more than half died due to vascular causes. Malignancy, heart failure, heavy drinking, preceding infection, type 1 diabetes, increasing age, and large-artery atherosclerosis causing the index stroke independently predicted 5-year mortality when adjusted for age, gender, relevant risk factors, stroke severity, and etiologic subtype. In sum, young adults with ischemic stroke have distinct demographic patterns and they frequently harbor traditional vascular risk factors. Etiology in the young is extremely diverse, but in as many as one-third the exact cause remains unknown. Silent brain infarcts and leukoaraiosis are not uncommon brain imaging findings in these patients and should not be overlooked due to their potential prognostic relevance. Outcomes in young adults with hemispheric ischemic stroke can safely be improved with intravenous thrombolysis. Furthermore, despite their overall low risk of death after ischemic stroke, several easily recognizable factors—of which most are modifiable—predict higher mortality in the long term in young adults.

Dynamic nature of proteins - interpretation of residual dipolar couplings

Protein conformations and dynamics can be studied by nuclear magnetic resonance spectroscopy using dilute liquid crystalline samples. This work clarifies the interpretation of residual dipolar coupling data yielded by the experiments. It was discovered that unfolded proteins without any additional structure beyond that of a mere polypeptide chain exhibit residual dipolar couplings. Also, it was found that molecular dynamics induce fluctuations in the molecular alignment and doing so affect residual dipolar couplings. The finding clarified the origins of low order parameter values observed earlier. The work required the development of new analytical and computational methods for the prediction of intrinsic residual dipolar coupling profiles for unfolded proteins. The presented characteristic chain model is able to reproduce the general trend of experimental residual dipolar couplings for denatured proteins. The details of experimental residual dipolar coupling profiles are beyond the analytical model, but improvements are proposed to achieve greater accuracy. A computational method for rapid prediction of unfolded protein residual dipolar couplings was also developed. Protein dynamics were shown to modulate the effective molecular alignment in a dilute liquid crystalline medium. The effects were investigated from experimental and molecular dynamics generated conformational ensembles of folded proteins. It was noted that dynamics induced alignment is significant especially for the interpretation of molecular dynamics in small, globular proteins. A method of correction was presented. Residual dipolar couplings offer an attractive possibility for the direct observation of protein conformational preferences and dynamics. The presented models and methods of analysis provide significant advances in the interpretation of residual dipolar coupling data from proteins.

Residual noise covariance for Planck low-resolution data analysis

Aims: Develop and validate tools to estimate residual noise covariance in Planck frequency maps. Quantify signal error effects and compare different techniques to produce low-resolution maps. Methods: We derive analytical estimates of covariance of the residual noise contained in low-resolution maps produced using a number of map-making approaches. We test these analytical predictions using Monte Carlo simulations and their impact on angular power spectrum estimation. We use simulations to quantify the level of signal errors incurred in different resolution downgrading schemes considered in this work. Results: We find an excellent agreement between the optimal residual noise covariance matrices and Monte Carlo noise maps. For destriping map-makers, the extent of agreement is dictated by the knee frequency of the correlated noise component and the chosen baseline offset length. The significance of signal striping is shown to be insignificant when properly dealt with. In map resolution downgrading, we find that a carefully selected window function is required to reduce aliasing to the sub-percent level at multipoles, ell > 2Nside, where Nside is the HEALPix resolution parameter. We show that sufficient characterization of the residual noise is unavoidable if one is to draw reliable contraints on large scale anisotropy. Conclusions: We have described how to compute the low-resolution maps, with a controlled sky signal level, and a reliable estimate of covariance of the residual noise. We have also presented a method to smooth the residual noise covariance matrices to describe the noise correlations in smoothed, bandwidth limited maps.