Metabolts is a software library for meta-learning research.Metabolts: A library for high-end research in Meta Learning.Motivation: Build efficient algorithms that surpass SOTA performance on existing CV benchmark datasets.Tree MAML & Hierarchical MT-NET (ongoing work).Additional Material: Code / Poster / Presentation (Progress) / Presentation (Final).The dataset for the same is prepared by making use of Tweepy. Furthermore, we make use of FasTText and Transformer models to build a text classifier for classifying between states in a particular country. Our system is based on the previous systems Carmen and Pigeo, and we overcome some of the drawbacks present in these systems while retaining the advantages of these systems. Herein, we present a system which makes use of the location metadata such as geo, place and profile tags and make use of user tweets to predict the structured location of a given Twitter user. In the past few years, the importance of geolocation systems have increased in health care applications, advertisments, natural disaster management, etc. Overview: Non-availability of well annotated and balanced datasets is considered as one of the major hurdles in analysing and extracting meaningful information from health-related tweets.Topic: Geolocation Extraction of Twitter users.Transformer Models for Classification on Health-Related Imbalanced Twitter Datasets.Performed extensive research in the fields of Fair Division, Voting and studied to hardness of various problems.Contributed a lemma to reduce the NP-Hard problem of winner determination for connected property of candidate graph for NSAV/NAV scoring functions to polynomial time solvable for a fixed size of voters.Topic: Multiwinner voting with Admissible Sets.Game Theory Lab, Indian Insitute of Technology, Gandhinagar.The task was to classify COVID-19 tweets as informative or not information. Participated in WNUT’20 shared task, where our system ranked 13 th all over the world.Implemented and improved on existing algorithms in literature to achieve SOTA performance for extraction of health-related information (eg., age, gender, location, symptoms, etc.) using Multi-task learning from multi-corpus twitter health dataset.Graciela Gonzalez-Hernandez, Associate Professor Topic: Multi-task learning on multi-corpus health dataset.Health Language Processing Lab, University of Pennsylvania.Additionally, incorporated Data Parallel & Distributed settings. Contributed to an open-source meta learnning library to generalize performance optimization achieved on MAML, ANIL, etc.Optimized system and model performance of a RNN based meta learner, thus achieving a speed up of 4x on training.Rekha Singhal, Principle Scientist & Head of Lab Topic: Accelerating Gradient Based Meta Learners.Computing Systems Lab, Tata Research & Innovation Labs, Mumbai.IIITN: Indian Institute of Information Technology, Nagpur. B.Tech., Electronics and Communication Engineering, IIITN.The list is by means not exhaustive, as I am always open to dabble into new areas □. Meta Learning / NLP / Algorithmic Game Theory / AI for Healthcare
0 Comments
Leave a Reply. |
AuthorWrite something about yourself. No need to be fancy, just an overview. ArchivesCategories |