Mike Gardner

Data mining solves tough semiconductor manufacturing problems

Quickly solving product yield and quality problems in a complex manufacturing process is becoming increasingly more difficult. The “low hanging fruit” has been plucked using process control, statistical analysis, and design of experiments which have established a solid base for a well tuned manufacturing process. However, the dynamic “higher-tier” problems coupled with quicker time to market expectations is making finding and resolving problems quickly an overwhelming task. These dynamic “higher tier” problems include: multi-factor & nonlinear interactions; intermittent problems; dynamically changing processes; installing new processes; multiple products; and, of course, the increasing volumes of data. Data mining technology can increase product yield and quality to the next higher level by quickly finding and solving these tougher problems. Case studies of semiconductor wafer manufacturing problems are presented. A combination of self-organizing neural networks and rule induction is used to identify the critical poor yield factors from normally collected wafer manufacturing data. Subsequent controlled experiments and process changes confirmed the solutions. Wafer yield problems were solved 10x faster than standard approaches; yield increases ranged from 3% to 15%; endangered customer product deliveries were saved. This approach is flexible and can be appropriate for a number of complex manufacturing processes

Using Machine Learning Techniques to Reduce Data Annotation Time

Manually annotating large databases in any domain is costly and time-consuming. We present a semi-automatic annotation tool for this purpose that uses Random Forests as bootstrapped classifiers. We describe an application of this tool on a large database of simulated driving data. The tool enables the user to verify automatically generated annotations, rather than annotating from scratch. This tool reduced the amount of time required to annotate one minute of video by a factor of six, down to approximately thirty-five seconds of annotation time per minute of video for a database of simulated driving data. The tool is limited in that its effectiveness is dependent upon the types of data collected, and the statistical boundaries between the different annotations.

Adaptive learning of semantic locations and routes

Adaptation of devices and applications based on contextual information has a great potential to enhance usability and mitigate the increasing complexity of mobile devices. An important topic in context-aware computing is to learn semantic locations and routes of mobile device users. Several batch methods have been proposed to learn these locations. However, such offline methods have very limited usefulness in practice. This paper describes an online adaptive approach to learn users semantic locations. The proposed method models users GPS data as a mixture of Gaussians, which is updated by an online estimation. The learned Gaussian mixture is then evaluated to determine which components most likely correspond to the important locations based on a priori probabilities. With learned semantic locations, we also propose a minimax criterion to discover users frequent transportation routes, which are modeled as sequences of GPS data. Finally, we describe an application of the proposed methods in a cell phone based automatic traffic advisory system.

A Context Aware Automatic Traffic Notification System for Cell Phones

Adaptation of devices and applications based on contextual information has a great potential to enhance usability and mitigate the increasing complexity of mobile devices. We have developed a context aware system that can automatically notify user of the traffic conditions by predicting their destination and route using context. In this paper, we describe the system in detail, including data collection, user interaction, context fusion, and learning and prediction of users destination and route.

Sensor Selection for Driving State Recognition

Driver activity recognition in the car cockpit is a necessary component for intelligent driver assistance systems. Since this has to be based on the sensor data stream available from the vehicle, an important question is what sensors are necessary and for which driver activities. We present results of a large-scale sensor selection study with naturalistic driving data looking at driving maneuver classification using ensemble methods.

Toward modeling and classification of naturalistic driving

Intelligent systems in automobiles need to be aware of the driving and driver context Available sensor data streams have to be modeled and monitored in order to do so. We describe a machine learning approach to accomplish this through collecting a large database of naturalistic driving data in a driving simulator. Preliminary experiments with a smaller dataset indicate successful modeling of naturalistic driving with hierarchical sequential models such as hidden Markov models.

Development of a Semi-Automatic Data Annotation Tool for Driving Data

Data-driven approaches to constructing context aware driver assistance systems require large annotated databases of automobile sensor data. Manually annotating such large databases is costly and time-consuming. We present a semi-automatic annotation tool for this purpose that uses random forests as bootstrapped classifiers. The tool significantly reduces the manual annotation effort by enabling the user to verify automatically generated annotations, rather than annotating from scratch