An interpretable method for automated classification of spoken transcripts and written text.

We investigate the differences between spoken language (in the form of radio show transcripts) and written language (Wikipedia articles) in the context of text classification. We present a novel, interpretable method for text classification, involving a linear classifier using a large set of n-gram features, and apply it to a newly generated data set with sentences originating either from spoken transcripts or written text. Our classifier reaches an accuracy less than 0.02 below that of a commonly used classifier (DistilBERT) based on deep neural networks (DNNs). Moreover, our classifier has an integrated measure of confidence, for assessing the reliability of a given classification. An online tool is provided for demonstrating our classifier, particularly its interpretable nature, which is a crucial feature in classification tasks involving high-stakes decision-making. We also study the capability of DistilBERT to carry out fill-in-the-blank tasks in either spoken or written text, and find it to perform similarly in both cases. Our main conclusion is that, with careful improvements, the performance gap between classical methods and DNN-based methods may be reduced significantly, such that the choice of classification method comes down to the need (if any) for interpretability.

Sleepy drivers on a slippery road: A pilot study using a driving simulator.

Sleepy drivers have problems with keeping the vehicle within the lines, and might often need to apply a sudden or hard corrective steering wheel movement. Such movements, if they occur while driving on a slippery road, might increase the risk of ending off road due to the unforgiving nature of slippery roads. We tested this hypothesis. Twelve young men participated in a driving simulator experiment with two counterbalanced conditions; dry versus slippery road × day (alert) versus night (sleepy) driving. The participants drove 52.5 km on a monotonous two-lane highway and rated their sleepiness seven times using the Karolinska Sleepiness Scale. Blink durations were extracted from an electrooculogram. The standard deviation of lateral position and the smoothness of steering events were measures of driving performance. Each outcome variable was analysed with mixed-effect models with road condition, time-of-day and time-on-task as predictors. The Karolinska Sleepiness Scale increased with time-on-task (p < 0.001) and was higher during night drives (p < 0.001), with a three-way interaction suggesting a small increased sleepiness with driving time at night with slippery road conditions (p = 0.012). Blink durations increased with time-on-task (p < 0.01) with an interaction between time-of-day and road condition (p = 0.040) such that physiological sleepiness was lower for sleep-deprived participants in demanding road conditions. The standard deviation of lateral position increased with time-on-task (p = 0.026); however, during night driving it was lower on a slippery road (p = 0.025). The results indicate that driving in demanding road condition (i.e. slippery road) might further exhaust already sleepy drivers, although this is not clearly reflected in driving performance.

Lidar-Camera Semi-Supervised Learning for Semantic Segmentation.

In this work, we investigated two issues: (1) How the fusion of lidar and camera data can improve semantic segmentation performance compared with the individual sensor modalities in a supervised learning context; and (2) How fusion can also be leveraged for semi-supervised learning in order to further improve performance and to adapt to new domains without requiring any additional labelled data. A comparative study was carried out by providing an experimental evaluation on networks trained in different setups using various scenarios from sunny days to rainy night scenes. The networks were tested for challenging, and less common, scenarios where cameras or lidars individually would not provide a reliable prediction. Our results suggest that semi-supervised learning and fusion techniques increase the overall performance of the network in challenging scenarios using less data annotations.

Alcohol ignition interlocks in all new vehicles: a broader perspective.

OBJECTIVE: To discuss the implications of widespread implementation of alcohol ignition interlocks. METHOD: We base our discussion on data from Finland including crash statistics and surveys collected from criminal justice professionals and general driving population. RESULTS: Alcohol ignition interlocks are an effective preventive measure against drunk driving when installed in the vehicles of convicted drunk drivers. However, once they are removed from the vehicles, drivers typically return to their habit of drinking and driving. Furthermore, for a number of reasons, the proportion of convicted drunk drivers that install an interlock in their vehicles is quite small. Therefore, many stakeholders believe that the solution to the drunk driving problem will come when interlocks become standard equipment in all new vehicles. However, drunk driving is a complex sociopsychological problem, and technology can rarely offer a solution to such complex problems. Consequently, many aspects of such interventions might be difficult to identify and include in cost-benefit analysis. CONCLUSION: We express caution about requiring an interlock as standard equipment in all new vehicles.

Effects of experience and electronic stability control on low friction collision avoidance in a truck driving simulator.

Two experiments were carried out in a moving-base simulator, in which truck drivers of varying experience levels encountered a rear-end collision scenario on a low-friction road surface, with and without an electronic stability control (ESC) system. In the first experiment, the drivers experienced one instance of the rear-end scenario unexpectedly, and then several instances of a version of the scenario adapted for repeated collision avoidance. In the second experiment, the unexpected rear-end scenario concluded a stretch of driving otherwise unrelated to the study presented here. Across both experiments, novice drivers were found to collide more often than experienced drivers in the unexpected scenario. This result was found to be attributable mainly to longer steering reaction times of the novice drivers, possibly caused by lower expectancy for steering avoidance. The paradigm for repeated collision avoidance was able to reproduce the type of steering avoidance situation for which critical losses of control were observed in the unexpected scenario and, here, ESC was found to reliably reduce skidding and control loss. However, it remains unclear to what extent the results regarding ESC benefits in repeated avoidance are generalisable to unexpected situations. The approach of collecting data by appending one unexpected scenario to the end of an otherwise unrelated experiment was found useful, albeit with some caveats.

Sleepy driving on the real road and in the simulator--A comparison.

Sleepiness has been identified as one of the most important factors contributing to road crashes. However, almost all work on the detailed changes in behavior and physiology leading up to sleep related crashes has been carried out in driving simulators. It is not clear, however, to what extent simulator results can be generalized to real driving. This study compared real driving with driving in a high fidelity, moving base, driving simulator with respect to driving performance, sleep related physiology (using electroencephalography and electrooculography) and subjective sleepiness during night and day driving for 10 participants. The real road was emulated in the simulator. The results show that the simulator was associated with higher levels of subjective and physiological sleepiness than real driving. However, both for real and simulated driving, the response to night driving appears to be rather similar for subjective sleepiness and sleep physiology. Lateral variability was more responsive to night driving in the simulator, while real driving at night involved a movement to the left in the lane and a reduction of speed, both of which effects were absent in the simulator. It was concluded that the relative validity of simulators is acceptable for many variables, but that in absolute terms simulators cause higher sleepiness levels than real driving. Thus, generalizations from simulators to real driving must be made with great caution.

A review of near-collision driver behavior models.

OBJECTIVE: This article provides a review of recent models of driver behavior in on-road collision situations. BACKGROUND: In efforts to improve traffic safety, computer simulation of accident situations holds promise as a valuable tool, for both academia and industry. However, to ensure the validity of simulations, models are needed that accurately capture near-crash driver behavior, as observed in real traffic or driving experiments. METHOD: Scientific articles were identified by a systematic approach, including extensive database searches. Criteria for inclusion were defined and applied, including the requirement that models should have been previously applied to simulate on-road collision avoidance behavior. Several selected models were implemented and tested in selected scenarios. RESULTS: The reviewed articles were grouped according to a rough taxonomy based on main emphasis, namely avoidance by braking, avoidance by steering, avoidance by a combination of braking and steering, effects of driver states and characteristics on avoidance, and simulation platforms. CONCLUSION: A large number of near-collision driver behavior models have been proposed. Validation using human driving data has often been limited, but exceptions exist. The research field appears fragmented, but simulation-based comparison indicates that there may be more similarity between models than what is apparent from the model equations. Further comparison of models is recommended. APPLICATION: This review provides traffic safety researchers with an overview of the field of driver models for collision situations. Specifically, researchers aiming to develop simulations of on-road collision accident situations can use this review to find suitable starting points for their work

The characteristics of sleepiness during real driving at night--a study of driving performance, physiology and subjective experience.

STUDY OBJECTIVES: Most studies of sleepy driving have been carried out in driving simulators. A few studies of real driving are available, but these have used only a few sleepiness indicators. The purpose of the present study was to characterize sleepiness in several indicators during real driving at night, compared with daytime driving. DESIGN: Participants drove 55 km (at 90 km/h) on a 9-m-wide rural highway in southern Sweden. Daytime driving started at 09:00 or 11:00 (2 groups) and night driving at 01:00 or 03:00 (balanced design). SETTING: Instrumented car on a real road in normal traffic. PARTICIPANTS: Eighteen participants drawn from the local driving license register. INTERVENTIONS: Daytime and nighttime drives. MEASUREMENT AND RESULTS: The vehicle was an instrumented car with video monitoring of the edge of the road and recording of the lateral position and speed. Electroencephalography and electrooculography were recorded, together with ratings of sleepiness every 5 minutes. Pronounced effects of night driving were seen for subjective sleepiness, electroencephalographic indicators of sleepiness, blink duration, and speed. Also, time on task showed significant effects for subjective sleepiness, blink duration, lane position, and speed. Sleepiness was highest toward the end of the nighttime drive. Night driving caused a leftward shift in lateral position and a reduction of speed. The latter two findings, as well as the overall pattern of sleepiness indicators, provide new insights into the effects of night driving. CONCLUSION: Night driving is associated with high levels of subjective, electrophysiologic, and behavioral sleepiness.

Reaction of sleepiness indicators to partial sleep deprivation, time of day and time on task in a driving simulator--the DROWSI project.

Studies of driving and sleepiness indicators have mainly focused on prior sleep reduction. The present study sought to identify sleepiness indicators responsive to several potential regulators of sleepiness: sleep loss, time of day (TOD) and time on task (TOT) during simulator driving. Thirteen subjects drove a high-fidelity moving base simulator in six 1-h sessions across a 24-h period, after normal sleep duration (8 h) and after partial sleep deprivation (PSD; 4 h). The results showed clear main effects of TOD (night) and TOT but not for PSD, although the latter strongly interacted with TOD. The most sensitive variable was subjective sleepiness, the standard deviation of lateral position (SDLAT) and measures of eye closure [duration, speed (slow), amplitude (low)]. Measures of electroencephalography and line crossings (LCs) showed only modest responses. For most variables individual differences vastly exceeded those of the fixed effects, except for subjective sleepiness and SDLAT. In a multiple regression analysis, SDLAT, amplitude/peak eye-lid closing velocity and blink duration predicted subjective sleepiness bouts with a sensitivity and specificity of about 70%, but were mutually redundant. The prediction of LCs gave considerably weaker, but similar results. In summary, SDLAT and eye closure variables could be candidates for use in sleepiness-monitoring devices. However, individual differences are considerable and there is need for research on how to identify and predict individual differences in susceptibility to sleepiness.

A survey of methods for classification of gene expression data using evolutionary algorithms.

The rapid increase in the quantity of available biologic data over the last decade, brought about by the introduction of massively parallel methods for gene expression measurements, has highlighted the need for more efficient computational techniques for analysis. This paper reviews the use of evolutionary algorithms (EAs) in connection with classification based on gene expression data matrices. Brief introductions to data classification methods and EAs are given, followed by a survey of studies dealing with the application of evolutionary algorithms to various (cancer related) data sets. The general conclusion, based on the published results surveyed here, is that EAs may constitute an efficient method for optimal gene selection, and can also help in reducing the size (number of features used) of classifiers. In many cases, the classification accuracy obtained using EAs, often in conjunction with other methods, represents a significant improvement over results obtained without the use of EAs. However, long-term, independent clinical follow-up studies will be essential to validate prognostic markers identified by the use of EA-based methods.

Effective dimensionality for principal component analysis of time series expression data.

Large-scale expression data are today measured for thousands of genes simultaneously. This development has been followed by an exploration of theoretical tools to get as much information out of these data as possible. Several groups have used principal component analysis (PCA) for this task. However, since this approach is data-driven, care must be taken in order not to analyze the noise instead of the data. As a strong warning towards uncritical use of the output from a PCA, we employ a newly developed procedure to judge the effective dimensionality of a specific data set. Although this data set is obtained during the development of rat central nervous system, our finding is a general property of noisy time series data. Based on knowledge of the noise-level for the data, we find that the effective number of dimensions that are meaningful to use in a PCA is much lower than what could be expected from the number of measurements. We attribute this fact both to effects of noise and the lack of independence of the expression levels. Finally, we explore the possibility to increase the dimensionality by performing more measurements within one time series, and conclude that this is not a fruitful approach.

Effective dimensionality of large-scale expression data using principal component analysis.

Large-scale expression data are today measured for thousands of genes simultaneously. This development is followed by an exploration of theoretical tools to get as much information out of these data as possible. One line is to try to extract the underlying regulatory network. The models used thus far, however, contain many parameters, and a careful investigation is necessary in order not to over-fit the models. We employ principal component analysis to show how, in the context of linear additive models, one can get a rough estimate of the effective dimensionality (the number of information-carrying dimensions) of large-scale gene expression datasets. We treat both the lack of independence of different measurements in a time series and the fact that that measurements are subject to some level of noise, both of which reduce the effective dimensionality and thereby constrain the complexity of models which can be built from the data.

Assessing the significance of consistently mis-regulated genes in cancer associated gene expression matrices.

MOTIVATION: The simplest level of statistical analysis of cancer associated gene expression matrices is aimed at finding consistently up- or down-regulated genes within a given set of tumor samples. Considering the high level of gene expression diversity detected in cancer, one needs to assess the probability that the consistent mis-regulation of a given gene is due to chance. Furthermore, it is important to determine the required sample number that will ensure the meaningful statistical analysis of massively parallel gene expression measurements. RESULTS: The probability of consistent mis-regulation is calculated in this paper for binarized gene expression data, using combinatorial considerations. For practical purposes, we also provide a set of accurate approximate formulas for determining the same probability in a computationally less intensive way. When the pool of mis-regulatable genes is restricted, the probability of consistent mis-regulation can be overestimated. We show, however, that this effect has little practical consequences for cancer associated gene expression measurements published in the literature. Finally, in order to aid experimental design, we have provided estimates on the required sample number that will ensure that the detected consistent mis-regulation is not due to chance. Our results suggest that less than 20 sufficiently diverse tumor samples may be enough to identify consistently mis-regulated genes in a statistically significant manner. AVAILABILITY: An implementation using Mathematica (tm) of the main equation of the paper, (4), is available at www.me.chalmers.se/~mwahde/bioinfo.html.