Search results
(1 - 3 of 3)
- Title
- ACTIVE LEARNING WITH RICH FEEDBACK
- Creator
- Sharma, Manali
- Date
- 2017, 2017-07
- Description
-
One of the goals of artificial intelligence is to build predictive models that can learn from examples and make predictions. Predictive models...
Show moreOne of the goals of artificial intelligence is to build predictive models that can learn from examples and make predictions. Predictive models are useful in many domains and applications such as predicting fraud in credit card transactions, predicting whether a patient has heart-disease, predicting whether an email is a spam, predicting crime, recognizing images, recognizing speech, and many more. Building predictive models often requires supervision from a human expert. Since there is a human in the loop, the supervision needs to be as resource-efficient as possible to save the human’s time, cost, and effort in providing supervision. One solution to make the supervision resource-efficient is active learning, in which the active learner interacts with the human to acquire supervision, usually in the form of labels, for a few selected examples to effectively learn a function that can be used to make predictions. In this thesis, I explore more intuitive and effective use of human supervision through richer interactions between the human expert and the learner, so that the human can understand the learner’s reasoning for querying examples, and provide information beyond just the labels for examples. Traditional active learning approaches select informative examples for labeling, but the human does not get to know why those examples are useful to the learner. While interacting with the learner to annotate examples, humans can provide rich feedback, such as provide their prior knowledge and understanding of the domain, explain certain characteristics of the data, suggest important attributes of the data, give rationales for why an example belongs to a certain category, and provide explanations by pointing out features that are indicative of certain labels. The challenge, however, is that traditional supervised learning algorithms can learn from labeled examples, but they are not equipped to readily absorb the rich feedback. In this thesis, we enable the learner to explain its reasons for selecting instances and devise novel methods to incorporate rich feedback from humans into the training of predictive models. Specifically, I build and evaluate four novel active learning frameworks to enrich the interactions between the human and learner. First, I introduce an active learning framework to reveal the learner’s perception of informative instances. Specifically, we enable the learner to provide its reasons for uncertainty on examples and utilize the learner’s perception of uncertainty to select better examples for training the predictive models. Second, I introduce a framework to enrich the interaction between the human and learner for document classification task. Specifically, we ask the human to annotate documents and provide rationales for their annotation by highlighting phrases that convinced them to choose a particular label for a document. Third, I introduce a framework to enrich the interaction between the human and learner for the aviation domain, where we ask subject matter experts to examine flights and provide rationales for why certain flights have safety concerns. Fourth, I introduce a framework to enrich the interaction between the human and learner for document classification task, where we ask humans to provide explanations for classification by highlighting phrases that reinforce their belief in the document’s label and striking-out phrases that weaken their belief in the document’s label. We show that enabling richer interactions between the human and learner and incorporating rich feedback into learning lead to more effective training of predictive models and better utilization of human supervision.
Ph.D. in Computer Science, July 2017
Show less
- Title
- LIGHTLY SUPERVISED MACHINE LEARNING FOR CLASSIFYING ONLINE SOCIAL DATA
- Creator
- Mohammady Ardehaly, Ehsan
- Date
- 2017, 2017-05
- Description
-
Classifying latent attributes of social media users has many applications in public health, politics, and marketing. For example, web-based...
Show moreClassifying latent attributes of social media users has many applications in public health, politics, and marketing. For example, web-based studies of public health require monthly estimates of the health status and demographics of users based on their public communications. Most existing approaches are based on supervised learning. Supervised learning requires human annotated labeled data, which can be expensive and many attributes such as health are hard to annotate at the user level. In this thesis, we investigate classification algorithms that use population statistical constraints such as demographics, names, polls, and social network followers to predict individual user attributes. For example, the racial makeup of counties is a source of light supervision came from the U.S. Census to train classification models. These statistics are usually easy to obtain, and a large amount of unlabeled data from social media sites (e.g. Twitter) are available. Learning from Label Proportions (LLP) is a lightly supervised approach when the training data is multiple sets of unlabeled samples and only label distributions of them are known. Because social media users are not a representative sample of the population and constraints are too noisy, using existing LLP models (e.g. linear models, label regularization) is insufficient. We develop several new LLP algorithms to extend LLP to deal with this bias, including bag selection and robust classification models. Also, we propose a scalable model to infer political sentiment from the high temporal big data, and estimate the daily conditional probability of different attributes as a supplement method to polls, for social scientists. Because, constraints are not often available in some domains (e.g. blogs), we propose a self-training algorithm to gradually adapt a classifier trained on social media to a different but similar field. We also extend our framework to deep learning and provide empirical results for demographic classification using the user profile image. Finally, when both textual and profile image are available for a user, we provide a co-training algorithm to iteratively improve both image and text classifications accuracy, and apply an ensemble method to achieve the highest precision.
Ph.D. in Computer Science, May 2017
Show less
- Title
- Machine Learning at the Bureau of Labor Statistics
- Creator
- Ellis, Robert, Kannan, Vinesh
- Date
- 2019-11-21
- Description
-
Vinesh Kannan (CS '19) shares his experiences working as a...
Show moreVinesh Kannan (CS '19) shares his experiences working as a data science fellow at the Bureau of Labor Statistics (BLS). Vinesh worked on the team that produces occupation and wage data used by policymakers, hiring staff, job seekers, and researchers across the country. He helped improve machine learning systems at the BLS: automatically identifying problematic training data and classifying rare jobs. Vinesh offers advice for students who may be interested in applying for the 2020 Civic Digital Fellowship, a program that recruits university students at all levels to spend a summer working on civic technology projects with various federal agencies.
Sponsorship: College of Science, Department of Computer Science, Department of Applied Mathematics, Machine Learning at IIT
Show less