In early 1999, VA Software launched a project to understand how the Internet development community had been able to produce software such as Linux, Apache and Samba that was generally developed faster and with higher quality than comparable commercially available alternatives [1,2,3,20]. Our goal was simple: determine how to make more software development projects successful.We discovered that successful Internet community projects employed a number of practices that were not well characterized by traditional software engineering methodologies. We now refer to those practices as Collaborative Software Development or CSD. Late in 1999 we developed the SourceForge platform to make it easy for even small software development projects to employ those practices, and in November of 1999 launched the SourceForge.net web site based on the SourceForge platform.The site was an overwhelming success, and in less than two years, grew to support more than 27,000 software development projects and over a quarter million software developers worldwide. SourceForge.net affords us an unequaled test bed for understanding CSD. In response to demand from companies seeking to enable CSD within their organizations, we announced a commercial version of the SourceForge platform, SourceForge Enterprise Edition, in August 2001.This paper describes the principles of CSD, the software development pain points those principles address, and our experience enabling CSD with the SourceForge platform.

Gazula and Kabuka (1995) describe a binary neural network which implements a nonparametric statistical classifier. However, they implement a kernel-based classifier rather than k-nearest-neighbors, as stated in their paper. The commenter states that some other aspects of their paper are not clear. One of the original authors replies to points made by the commenter.

Image texture can be defined as a local two-dimensional random field. The Gauss Markov random field (GMRF) and grey level co-occurrence (GLC) algorithms compute features from models of this random field. However, the GMRF and GLC algorithms capture only second-order interactions between pixels. We describe an algorithm which models texture as a local two-dimensional random field and captures high-order interactions.