Research Projects
Balancing data in healthcare: A review of ensemble diversity
Data imbalance has continuously evolved to plague the field of applied machine learning. Critical applications such as healthcare often deal with highly imbalanced datasets where the minority class is usually of significant importance. Generic ML models fail to perform well on the minority class. With the introduction of SMOTE and subsequent variants, oversampling algorithms became one of the most widely adopted solutions for imbalanced datasets. This work reviews a subset of SMOTE variants and the performance improvements they provide on commonly used imbalanced datasets. These reviews are evaluated from the perspective of ensemble diversity using kappa error diagrams, the importance of which is explained. A new and intuitive metric is also proposed over kappa-error diagrams to evaluate the diversity of an ensemble.
Publisher: Expert Systems With Applications | Journal (In review) | 2024
MetaCirc: A Meta-Learning Approach for Statistical Leakage Estimation Improvement in Digital Circuits
Aggressive scaling down of transistor dimensions has made process-aware circuit modeling a crucial task. Achieving accurate circuit modeling requires lengthy and resource-intensive simulations. Machine Learning-based surrogate models, offering computational efficiency and speed, are viable alternatives to traditional simulators. This work introduces a meta-learning approach designed to accurately capture process-induced variations in the leakage power of VLSI circuits. We use a meta-learning architecture that leverages a learnable neural network to choose between pre-trained regressors through a soft-choice process for a particular data point.
Publisher: IEEE ISCAS | Conference | 2024
View Publication | Video | Showcase
RiQualis: A portable and accessible smartphone based Rice Quality Analysis tool
Rice (Oryza Sativa) is a staple food for millions of people across the globe and is known to supply about 20% of calories to the population of the world each day. It is one of the most widely consumed food crops in Asia where it provides over 50% of the calorific supply. The quality analysis of such a popular and widely consumed crop is of paramount importance as many people are dependent on rice to meet their nutrition requirements. However a large quantity of the quality analysis methods used in the industry call for manual inspection that leads to inaccurate and time consuming analysis. By introducing a method to effectively utilize smartphone technology, we improve the accessibility and portability of existing methods such as so that farmers having budget smartphones can also utilize this technology for their benefits. The tool achieved an accuracy of about 96% on the test set and the algorithms were able to detect and extract the physical features of the rice grains with an accuracy of pm 0.134 mm. We use the smartphone as a sensor to capture images and then perform analysis on the image. Our proposed method allows people with limited access to technology such as farmers to perform accurate analysis of rice grains by making use of available gadgets such as smartphones. The method can also be easily extended for other edible food grains too.
Publisher: NA | Journal (Draft) | 2022