-
The aim of this project is to develop a fully functional GUI application which checks input text for grammatical and spelling errors. We have used N-grams model for this task. The data used were Gutenberg and web scraped data. Additional features include speech input (made possible from Google Speech API) and Spanish language support.
-
Automatic Speech Recognition is a widely explored field with its early systems based on Hidden Markov Models (HMMs) - Gaussian Mixture Model (GMM), but with the advent of neural networks, hybrid systems using HMM - DNN (Deep Neural Network) occupied this space. Later when the Connectionist Temporal Classification (CTC) Loss came about, end-to-end systems started to show promising results. This project is an attempt to obtain state of the art results for Australian English audio samples.
-
The focus of this project is the Variational Inference (VI) method and some of its applications. In particular, mean-field variational inference is used to find an estimate of the posterior distribution for each of the latent variables. This goal is obtained by decreasing Kullback-Leibler divergence as a measure of distance between the posterior distribution of the latent variables and a candidate distribution from the mean field of distributions.We begin by explaining the idea behind the Expectation-Maximization algorithm, and later we draw the analogy between the EM and the VI methods. Later on, we implement the VI on the univariate and multivariate Gaussian mixture models. In the end, we try to segment images based on their color map.
-
Image compression has an important role in data transfer and storage, especially due to the data explosion that is increasing significantly faster than Moore’s Law.[1] It is a challenging task since there are highly complex unknown correlations between the pixels, as a result, it is hard to find and recover them. We want to find a well-compressed representation for images and, design and test networks that are able to recover it successfully in a lossless or lossy way.
-
Neural networks are growing in stature day by day.Neural networks have been hugely successful in approximating fairly complex models.This project aims to implement a three layer network in C to classify the digits in MNIST data. First the baseline serial code was implemented from scratch. Then different optimization techniques such as loop unrolling, Intel intrinsics, OpenMP and CUDA were implemented. The performance of the optimized versions are compared with the serial baseline code and the results are presented.