Paper Reading #4- Gestalt: integrated support for implementation and analysis in machine learning

Title: Gestalt: integrated support for implementation and analysis in machine learning.
Reference Information:
Kayur Patel, Naomi Bancroft, Steven M. Drucker, James Fogarty,  Andrew J. Ko, and James Landay. "Gestalt: integrated support for implementation and analysis in machine learning". UIST '10 Proceedings of the 23nd annual ACM symposium on User interface software and technology ACM New York, NY, USA ©2010 ISBN: 978-1-4503-0271-5.
Author Bios:
Kayur Patel is a Ph.D student (in Computer Science) at the University of Washington. Part of the DuB group advised by James Fogarty and James Landay. His work has been funded at points by either Microsoft or the government.
Naomi Bancroft is an undergraduate student at the University of Washington studying Computer Science. After she graduates she will go work for Google.
Steven M. Drucker is an affiliate professor at the University of Washington. He is also a Principal Researcher and manager of the VUE group in VIBE at Microsoft Research.
James Fogarty is an assistant professor in the Computer Science department at the University of Washington. He is also involved in the DuB group.
Andrew J. Ko is an assistant professor in the information school at the University of Washington. He has traveled around the country giving special lectures on HCI topics. He is also a member of the group dub.
James Landay is a professor in the Computer Science department at the University of Washington. He is a founder of the dub group. He was previously the director of Intel labs in Seattle.
Summary:

• Hypothesis: A development environment for machine learning that is NOT domain-specific (aka a general purpose environment) is entirely feasible to create and implement.
• Methods: To prove their hypothesis, the authors set out to achieve two main goals: being able to implement a classification pipeline and being able to analyze the data as it moves through said pipeline. They decided to test the implementations for these goals by applying methods for solving two problems- sentiment analysis (categorizing text) and gesture recognition. The authors saw that in order to achieve a general purpose supporting environment, they needed to explicitly define many specific steps and structures to the user. Gestalt uses relational tables addressing the entire pipeline to effectively manage all of their general purpose data. Because they use a single relational table, data does not need to be converted at all and users don't have to switch between tools for editing. To support many different kinds of data, Gestalt uses aggregated visualizations along the pipeline, so that the data is all connected. The authors recruited 8 volunteers who had at least taken one python class and a machine learning algorithms course to test Gestalt vs a "baseline" software that served a similar purpose. The study was whether or not the students could identify and fix the injected bugs in some trials faster and more efficiently with Gestalt or with baseline.
• Results: Participants unanimously preferred Gestalt and were able to find and fix more bugs using Gestalt than using the baseline. After analyzing the study data, the authors found that the users spent more time analyzing rather than implementing in Gestalt and vice versa in baseline (which is preferrable). The students also were able to analyze the data with many more views then with baseline.
• Contents: In this paper, the authors sought to show off their created work called Gestalt. They tested it out in practical situations with typical users other than themselves by recruiting volunteers with relevant experience and who were competant enough to use Gestalt to come in and be a part of their experiement to measure the success of their creation. After each experiment was conducted, results were gathered. The whole point of this creation was to build on machine learning mechanisms in a general sort of way (one that is domain-independent).

Discussion:
I'll be honest, this paper was sort of confusing. I'm not competant enough in machine learning or pipelining to be able to follow some of the arguments or descriptions that the authors were making. Their creation was definitely successful and their goals of the project were achieved. This is evident because of the feedback from the users. I definitely admire these kinds of projects and these kinds of people and their intelligence. As far as the future goes for this sort of technology, I can see improvements being made on Gestalt and allowing a greater degree of machine learning to be capable. I am convinced by their work not because I followed it 100%, but because of the outstanding satisfaction that the users had with Gestalt.