Multiple Instance Learning

We consider the problem of training discriminative structured output predictors, such as conditional random fields (CRFs) and structured support vector machines (SSVMs). A generalized loss function is introduced, which jointly maximizes the entropy and the margin of the solution. The CRF and SSVM emerge as special cases of our framework. The probabilistic interpretation of large margin methods reveals insights about margin and slack rescaling. Furthermore, we derive the corresponding extensions for latent variable models ...

We present an efficient Hough transform for automatic detection of cylinders in point clouds. As cylinders are one of the most frequently used primitives for industrial design, automatic and robust methods for their detection and fitting are essential for reverse engineering from point clouds. The current methods employ automatic segmentation followed by geometric fitting, which requires a lot of manual interaction during modelling. Although Hough transform can be used for automatic detection of cylinders, the required 5D Hough space has a prohibitively high time and space complexity for most practical applications. We address this problem in this paper and present a sequential Hough transform for automatic detection of cylinders in point clouds. Our algorithm consists of two sequential steps of low dimensional Hough trans- forms. The first step, called Orientation Estimation, uses the Gaussian sphere of the input data and performs a 2D Hough Transform for finding strong hypotheses ...

Artificial intelligence in combination with modern technologies including medical screening devices has the potential to deliver better management services to deal with chronic diseases with higher accuracy, efficiency, and satisfaction. With the recent evolution in digitized data acquisition, computer vision and machine learning, AI solutions are spreading into areas which were previously examined by well-trained clinicians. Early diagnosis of diabetic retinopathy (DR) and foot ulcers (DFU) occurrence through image analysis is in high demand as many individuals are left without any supervision due to the limited resources such as trained clinicians or suitable equipment especially, in rural areas. Furthermore, the existing system will become even more insufficient as the number of people with diabetes increases. In this research paper, we propose a prototype that involves an autonomous system called an Intelligent Diabetic Assistant (IDA), which decides the diagnosis and the treatment prioritization depending upon the observations appeared in the screen. The IDA consists of knowledge-based modules for severity level-based classification, clinical decision support and near real- time foot ulcer detection and boundary screening. We use the System Usability Scale (SUS) in terms of performance; learn ability, and satisfaction to measure the usability of the IDA. The mean SUS score was 88.5, demonstrating good but not exceptional system usability. We perform our experiments with clinicians who have been involved in diabetic care.

Database integration provides integrated access to multiple data sources. Database integration has two main activities: schema integration (forming a global view of the data contents available in the sources) and data integration (transforming source data into a uniform format). This paper focuses on automating the aspect of data integration known as entity identification using data mining techniques. Once a global database is formed of all the transformed source data, there may be multiple instances of the same entity, with different values for the global attributes, and no global identifier to simplify the process of entity identification. We implement decision trees and k-NN as classification techniques, and we introduce a preprocessing step to cluster the data using conceptual hierarchies. We conduct a performance study using a small testbed and varying parameters such as training set size and number of unique entities to study processing speed and accuracy tradeoffs. We find th...

With the phenomenal growth of the Web resources, to construct ontologies by using existing resources structured in the Web has gotten more and more attention. Previous studies for constructing ontologies from the Web have not carefully considered all the semantic features of the Web documents. Hereby it is difficult to correctly construct ontology elements from the Web documents that are increasing daily. The machine learning methods play an important role in automatic constructing of the Web ontology. Bootstrapping technique is a semi-supervised learning method that can automatically generate many terms from the few seed terms entered by human. This paper proposes bootstrapping method that can automatically construct instances and data type properties of the Web ontology, taking proper noun as semantic core element of the Web table. Experimental result shows that proposed method can rapidly and effectually construct instances and its properties of the Web ontology.

Object descriptions used for 3D segmentation by deformable models and for statistical characterization of 3D object classes benefit from having intrinsic correspondences over deformation of the objects or multiple instances in the same object class. These correspondences apply over a variety of spatial scale levels and consequently lead to efficient segmentation and probability distributions of geometry that are trainable with

An ability to adjust to changing environments and un- foreseen circumstances is likely to be an important com- ponent of a successful autonomous space robot. This pa- per shows how to augment reinforcement learning algo- rithms with a method for automatically discovering cer- tain types of subgoals online. By creating useful new subgoals while learning, the agent is able to

We extend live sequence charts (LSCs), a highly expressive variant of sequence diagrams, and provide the extension with an executable semantics. The extension involves support for instances that can bind to multiple objects and symbolic variables that can bind to arbitrary values. The result is a powerful executable language for expressing behavioral requirements on the level of inter-object interaction. The extension is implemented in full in our play-engine tool, with which one can execute the requirements directly without the need to build or synthesize an intra-object system model. It seems that in addition to many advantages in testing and requirements engineering, for some kinds of systems this could lead to the requirements actually serving as the final implementation.