Publications

Objective: COLT aims to reach high sparsity levels in neural networks while preserving accuracy, challenging the existing benchmarks set by the Lottery Ticket Hypothesis (LTH).

Method: In the COLT method, a dataset is partitioned into N segments, and distinct models are trained on each segment, maintaining identical initial weights. For subsequent pruning iterations, only the intersecting pruned weights from these N models are retained.

Outcome: Unlike LTH, COLT achieves optimal 'winning lottery tickets' (efficiently pruned networks) in fewer iterations, thereby cutting down on computational expenses. Through successful trials on Cifar-10, Cifar-100, tinyand TinyImageNet, COLT not only demonstrates superior performance over current leading methods but also shows remarkable dataset transferability, enhancing both the convergence and compression aspects of neural network pruning.

Objective: Mitigate bias in handwritten OCR models for morphologically rich languages.

Method: Trained a selected set of cross-lingual balanced printed graphemes (generated) using cross-entropy loss and extracted the knowledge from it.

Outcome: Applied this knowledge to an OCR model trained on word-level handwritten data with CTC loss to enhance the performance of minor classes.

Objective: To address and reduce the high carbon footprint associated with knowledge distillation in deep learning models.

Method: Implementing a stochastic approach for selecting hyperparameters, aimed at minimizing energy consumption and CO2 emissions.

Outcome: Introducing and utilizing new metrics for assessing environmental costs (like GFLOPS, energy usage, and CO2 equivalents) in deep learning and promoting eco-friendly AI practices.

Objective: To distinguish intentional from unintentional opioid use through the application of machine learning and deep learning methods.

Method: Utilized structured and unstructured data from the MIMIC-III database, analyzing 455 patient cases, and employed knowledge distillation techniques to transfer insights from structured to unstructured datasets.

Outcome: Achieved 95% accuracy in identifying intentional users and 64% for accidental users in tests, with a further refined knowledge-based test accuracy of 76.44% after integrating both models. The study also offers new analytical insights into opioid patient profiles.

Objective: To surpass current state-of-the-art (SOTA) pretrained Bengali Automatic Speech Recognition (ASR) models.

Method: Finetuning a pretrained wav2vec2 model on the Common Voice dataset, coupled with the integration of an n-gram language model as a post-processor.

Outcome: Conducted various experiments and hyperparameter tuning to develop a more robust Bangla ASR model, achieving superior performance compared to existing models.

Objective: Mitigate bias in handwritten OCR models for morphologically rich languages.

Method: Trained a selected set of balanced printed graphemes (generated) using cross-entropy loss and extracted the knowledge from it.

Application: Applied this knowledge to an OCR model trained on word-level handwritten data with CTC loss to enhance the performance of minor classes.

FACT:  Drawing inspiration from the rich heritage of Bengali literature and the legendary mathematician Bhaskar II, who poignantly named his mathematical treatise "Lilavati" after his daughter, combined with the emotional depth in Bangladeshi author Humayun Ahmed's novel and personal story, the name "Lilaboti" was chosen for our paper. 

Objective: To monitor current air pollution levels and accurately predict future environmental pollution levels.

Method: Utilization of the ARIMA model, implemented via the Statsmodels package, for forecasting the levels of various pollutants.

Outcome: This system serves as a tool for both real-time pollution detection and future pollution level prediction, enhancing environmental monitoring capabilities.