2023

1. A Comparison of LSTM and GRU for Bengali Speech-to-Text Transformation

   Nusrat Jahan, Zakia Sultana, Fahim Chowdhury, and 4 more authors

   In Proceedings of the 2023 International Conference on Advances in Computing Research (ACR’23), 2023

   Abs

   This paper represents an approach to speech-to-text conversion in the Bengali language. In this area, we have found most of the methodologies were focused on other languages rather than Bengali. We started with a novel dataset of 56 unique words from 160 individual subjects was prepared. Then in this paper, we illustrate the approach to increasing accuracy in a speech-to-text over the Bengali language where initially we started with Gated Recurrent Unit(GRU) and Long short-term memory (LSTM) algorithms. During further observation, we found that the output of the GRU failed to give any stable output. So, we moved completely to the LSTM algorithm where we achieved 90% accuracy on an unexplored dataset. Voices of several demographic populations and noises were used to validate the model. In the testing phase, we tried a variety of classes based on their length, complexity, noise, and gender variant. Moreover, we expect that this research will help to develop a real-time Bengali speak-to-text recognition model.

2021

1. Benchmarking and Selecting Optimal Diabetic Retinopathy Detecting Machine Learning Model using Entropy and TOPSIS Method

   Sajjad Ahmed, Shaily Roy, and Golam Rabiul Alam

   In 2021 International Conference on Electrical, Computer, Communications and Mechatronics Engineering (ICECCME), 2021

   Abs

   Diabetic Retinopathy(DR) is a consequence of Diabetes Mellitus that manifests itself in the alteration of vascular tissue. When an alteration in the correct blood supply occurs, lesions such as micro-aneurysms, haemorrhages and exudates appear. A certain number of Machine Learning(ML) models are available to detect DR. Hence, a benchmarking method of these DR detection ML models can help to choose the best model based on some selected criteria. In this study, after pre-processing DR dataset, inception-v3 is used to extract the features from the images. Based on that, eight models are trained for detecting DR which are evaluated using ten metrics. Using those metrics as criteria, a decision matrix(DM) is formulated. Consequently, with respect to the weighted evaluation matrix, a Multi-Criteria Decision-Making(MCDM) method called Technique for Order Preference by Similarity to Ideal Solution (TOPSIS) is assigned to benchmark the DR detection models where Entropy is used to calculate the individual weight of these criteria. The result shows that with the integration of Entropy and TOPSIS, the evaluation and determining problems associated with DR detection methods can be resolved. The AdaBoost classifier is selected as the best detecting model for DR detection which satisfies all the criteria in the best possible way for the data studied. As demonstrated here, the weights are generated by Entropy method which can be replaced by customized weights as needed.

2019

1. Classification of Categorical Objects in Ventral Temporal Cortex using fMRI Data

   Sajjad Ahmed, and Mohammad Zavid Parvez

   In TENCON 2019 - 2019 IEEE Region 10 Conference (TENCON), 2019

   Abs

   Functional Magnetic Resonance Imaging(fMRI) is one of the best neuroimaging techniques which helps to understand the activity of the human brain. With the help of recent advancement in the field of machine learning algorithms in terms of pattern recognition, now it is possible to extract in-depth information about brain activity by analyzing fMRI data. In this paper, we have shown the analysis of the data of a particular part of the human brain called Ventral Temporal Cortex. The dataset contains the fMRI data of the subjects while viewing grey-scale image different categories of objects such as cat, chair, etc. We have applied the machine learning algorithms on the extracted feature set from fMRI data to classify the objects that the subject is viewing. Here, we have emphasized on hyper-parameter tuning for the classifiers. Among the classifiers, we have found that the performance of Support Vector Machine (i.e., 96.92%) and k-nearest neighbor classifier(i.e., 96.90%) is quite persistent and have better accuracy. The further application of this research may motivate to develop brain-computer interface (BCI) based solutions.

2. Pattern recognition with Quantum Support Vector Machine (QSVM) on near term quantum processors.

   Sajjad Ahmed

   Brac University, 2019

   Abs

   Machine Learning is boosting up the advancement in the eld of Arti cial Intelligence these days. However, almost every machine learning algorithm contains an optimization problem to solve. Inspired by quantum mechanics, quantum computing is quite a promising approach to solve high complexity optimization problems signi cantly faster and more e cient than classical computers. In this paper, we have worked on a very fundamental supervised learning problem. First, we discuss an approach to map the classical feature points on a quantum computer. Then we propose a Quantum Support Vector Machine(QSVM) model that runs on near term superconducting quantum processors. We show that using quantum optimization it is possible to train a discriminative SVM model that is capable of recognising patterns.