Projects
Genotype Imputation using bi-directional Recurrent Neural Network
Abstract
With the advancements in high-throughput single-nucleotide polymorphism (SNP) genotyping platforms, there has been an exponential increase in the number of measured genotypes. These advancements have enabled us to understand multi-gene interactions and also develop genetic models that more closely represent the polygenic nature of common disease risk. Studies have shown that the compounded effects of even small amounts of missing data on such analysis is considerable. In this work, we present a bidirectional recurrent neural network based method for imputing missing SNP genotype data. We perform our experiments on a subset of the 1000genomes dataset and our Bi-RNN method accurately predicts around 96% of the missing genotypes on average. We benchmark our results against the MENDEL imputation method and demonstrate that our bi-RNN based imputation method does as well other available methods such as MENDEL even with a vanilla architecture.
Yelp Restaurant Photo Classification using Convolutional Neural Network
Abstract
In this age of food selfies and photo-centric social storytelling, it may be no surprise to hear that Yelp’s users upload an enormous amount of photos every day alongside their written reviews. Every time a user uploads a photo of a restaurant, the person tags some attributes along with the image that closely describe the restaurant. For example, a user might upload a few photos of a restaurant and tag these images with labels such as ’restaurant is expensive”, ’good for kids’ etc. However till now, this tagging has to be done manually by the user. It would be useful to have a model that automatically recommends these tags for these restaurants based on the photos uploaded by the user. Our project is aimed towards this task of building a system that predicts the attribute labels of restaurants based on user-submitted photos.
Predicting and Detecting Freezing of Gait with Recurrent Neural Network / LSTM
Abstract
Freezing of Gait is a brief episodic disturbance of movement occurring in patients with advanced Parkinson’s disease. The occurrence is sudden, accompanied by falls and requires intensive patient monitoring. With the advent of miniature wearable sensors, we can capture the time series kinematics of limbs and perform data mining to discover pathologic patterns characteristic to FoG episodes. This can be combined with mobile based cueing systems to provide context- aware feedback to patient about the incidence of an episode.
In our work we proposed a novel algorithm for predicting the freezing of gait in Parkinson’s patients using an LSTM recurrent neural network. We went beyond state of the art for gait time series classification and use RNNs for ordered acceleration sequences. We also exploited a global training paradigm which trains the classifier with features of all patients and evaluate performance on a disjoint set of patients. Finally, we compared the performance of time-domain and frequency domain features for decision trees and LSTM for the prediction problem in patient dependent and independent studies.
Speaker Verification in Noisy Environments
Abstract
Speech allows humans to decipher important information about the speaker. One of them is personal identity. From voice itself, we are able to identify the person we are talking to over phone. Though it seems easy for humans, it has been very difficult for a machine to achieve the accuracy specially in a noisy environment. With wide applications in the field of security, automatic speaker recognition is currently an active area of research in speech processing. In this project we present a comparison of various state-of-the-art algorithms available for speaker verification. Also, we propose a new method using GFCCs coupled with pre and post processing of the speech signal for noise removal which gives better results in a noisy environment.