Research Projects

This project addresses the computational inefficiency of AI models that use various machine learning algorithms that rely on large datasets and many unknown parameters by utilizing Coresets. Coresets are weighted subsamples of complete datasets with provable theoretical guarantees, ensuring that models trained on these subsamples are good approximations of those trained on full datasets. This significantly improves training time and resource efficiency. Coresets have been studied for various machine learning problems widely used in AI models. There are three broad types of coresets: 1) 'Non-Deterministic coreset' ensures a lossy data summarization. A trained model on these coresets is approximately close to the optimal model trained on the full dataset with a high probability. So, these coresets suffer from a small non-zero failure probability. 2) 'Deterministic Coresets' are similar to the non-deterministic coreset, ensuring a lossy data summarization. However, the model approximation is guaranteed here. So, these coresets do not suffer from a failure probability. 3) 'Accurate Coresets' are also deterministic coresets. However, it ensures a lossless data summarization. Here, the model trained on the coreset is guaranteed equal to that trained on the full dataset with probability 1. The primary objectives of this research are to design new frameworks for deterministic (including accurate) coreset and to conduct thorough and detailed theoretical analysis to prove their correctness. A critical theoretical question we aim to address is to reduce the gap between the upper bound of the coreset from an algorithm and the existence of coreset's lower bound. The theoretical understanding of our frameworks will help answer this question in general. Another significant challenge for any framework is to demonstrate the existence of an efficient coreset for any given dataset, ensuring that the coreset size is significantly smaller than the full dataset. Further, to showcase the actual potential of our frameworks, we will design efficient (lightweight) coreset construction algorithms and test them on extensive, large, real datasets. The framework and analysis techniques will address many theoretical questions for any application with tailored loss functions. This will be further be useful in designing lightweight coresets for the problem. Coresets will be very useful in practice as: 1) It helps in reducing the resources required for large model/data training. 2) It is very efficient in realizing feature drift in a streaming dataset and detecting outliers. 3) Last but not least, it reduces the carbon footprint during the model training process, which is a step towards sustainable AI.