Types of Projects
Project Details
List of projects that presented with poster and demos.
-
Keystroke based Biometric Verification
Summary: We use keystroke pattern of an individual for the biometric authentication method. The keyboard stroke patterns of a user are continuously monitored, the flight and dwell time for each key are used, and the pattern of these extracted features is checked against the individual’s unique typing pattern to provide a continuous risk rating.
-
Toxic Comment Classification
Summary: Our study is basically focused on identification/classification of toxicity in a given text into multiple labels using CNNs and compare the results with other classifiers like SVM, Naïve Bayes and Random Forests. Our results indicate that CNNs outperform the others.
-
Medical Diagnosis through Retinal Fundus Images
Summary: A solution is established to treat blindness retinal disease based on deep learning framework. This solution utilizes optical coherence tomography (OCT) images of retinal layers and trains a neural network with scratch. A transfer learning technique has also been applied on image data and data augmentation has also been done. Lastly, significance of segmentation channels with corresponding input images have also been analysed.
-
Personalized Medicine: Redefining Cancer Treatment
Summary: We have developed a deep learning-based system that can mimic the work of a clinical pathologist by distinguishing between different types of genetic mutations, based on the associated clinical evidence. Our approach is tested on Memorial Sloan Kettering Cancer Centers’ (MSKCC) expert annotated dataset, which contains a list of 3500 genetic mutations along with the associated scientific literature that was used to assign them 9 different labels.
-
Classification of Sleep Disorder and Sleep Stages using Machine Learning Techniques
Summary: Polysomnographic is a multi-parametric test used in the study of sleep and as a diagnostic tool in sleep medicine. This works aims to provide a multi-level classification technique to automatically identify sleep stages and sleep related disorder in an effort to assist physicians in the diagnosis and treatment of sleep disorder. Four different classifiers including SVM and KNN have been trained using cross validation. Optimal features have been selected using forward feature selection.
-
Gunshot Detection
Summary: we have set a gunshot detection system with emphasis on quick response, robustness to external disturbances, and less runtime computations. For the achievement of these objectives we have utilized many comprehensive machine learning techniques, including support vector machines with various kernels, neural networks, convolutional neural networks, autoencoders and anomaly detection. By extracting a large set of features, training and testing on extensive dataset we were able to achieve best results in training and testing accuracy and in real time as well.
-
Predicting Fault Regime of Operation of Train Car [Poster] [Report]
Summary: Fault prediction in a complex mechanical system is one of the interesting topics for researchers. The goal of our project is to predict faulty regimes of a train car. The system we considered has accelerometers mounted on several parts and the dataset contains training and testing experiments. With the help of simplified physics based model we deduced 21 subsystems to detect any fault. We implemented Principle Component Analysis (PCA) anomaly detection by extracting data patterns from all training experiments and then each testing experiment has been classified either healthy or unhealthy. In the case of unhealthy experiment, faulty subsystem is detected.
-
Compressed Sensing MRI Reconstruction
Summary: We proposed a practical, software-only framework, based on deep learning, for accelerating MRI scan time allows maintaining good quality imaging with CS-MRI. This is accomplished by partial MRI sampling, while using an adversarial neural network to estimate the missing samples. The inter-play between the generator and the discriminator networks enables the introduction of an adversarial cost in addition to a fidelity loss used for optimizing the peak signal-to-noise ratio (PSNR).
-
Sales Forecasting on Walmart Data using Deep Learning
Summary: In this project, we aim to predict unit sales amount of each item in grocery stores based on their historical data, item/store information and oil price etc. For this purpose, we have employed simple MLP (Multi-Layer Perceptron) as baseline and build an LSTM model for accurate sales price forecasting. Our proposed model has a weighted MSE test loss of 0.529 comparing to the 0.85 of baseline.
-
Automatic Urdu Speech Recognition System
Summary: The main objective of our project is to design and implement a system which can recognize Urdu language from audio input. We hope that our Automatic Speech Recognition System (ASR) for Urdu will be an important stepping stone in the development of future systems which will allow the users to interact with them using Urdu speech.
-
Bitcoin Price Prediction using Machine Learning Techniques
Summary: Bitcoin is the first decentralized digital currency. This cryptocurrency was created in 2009 but it became extremely popular in 2017. This project is concerned with predicting the price of Bitcoin using machine learning. The goal is to ascertain with what accuracy can the direction of Bitcoin price in USD can be predicted. The price data is sourced from the Bitcoin Price Index . The task is achieved with varying degrees of success through the implementation of a Long Short Term Memory (LSTM) network.
-
Machine Learning approach to detect Industrial Plant Faults: PHM15 Data Challenge
Summary: Plant fault detection is a primary concern in plant operations and is a key variable in important statistics as productivity, profitability and efficiency. An efficient mechanism for fault detection translates in improvement of key statistics and as such warrants sizeable investments. This project is a PHM-society data challenge. In this project ML techniques are applied to plant data to assess the start time, end time and type of the fault for plants.
-
General Audio Tagging System
Summary: In Audio Tagging system we made feature vectors with Mel-frequency cepstral coefficients(MFCC)s and zero cross rating. Apply PCA on feature vectors for dimensionality reduction. A multiclass SVM audio classifier is designed to classify sounds in 41 classes. The results of experiments show that audio classification system classify audio signals with an average accuracy of 72%
-
Cricket Win Prediction
Summary: T20 is the shortest format of cricket and essentially the most exciting one. A T-20 game can tilt towards either side within a span of 3, 4 overs. This makes the outcome prediction of T20 match a real challenging problem. To overcome this challenge, we used team and player level statistics coupled with the score and wickets information at different intervals in 2 nd innings, in our machine learning model. Understandably, best results were achieved on predicting the outcome close to the end of 2 nd innings. Logistic Regression and SVM provided best results with overall accuracy of 84%.
-
Object detection using Oriented Response Networks
Summary: CNNs are translation invariant but not rotation invariant. To make CNNs rotation invariant, ORNs are designed that can detect rotated images with high accuracy and less number of training parameters.
-
Domain Adaptation of Semantic Segmentation
Summary: Domain Adaptation is one of the major problems in Semantic Segmentation where each pixel is labeled as a separate class. A solution is devised in-order to adapt the domain during semantic segmentation. The devised solution tries to adapt the target domain along with source domain using GANs approach. An experiment with fine tuning is also performed. The overall performance of the system is improved on target domain after adaptation on Cityscapes and WAD dataset respectively.
-
Urdu Speech to Text Using Transfer Learning
Summary: Traditional automatic speech recognition system was composed of acoustic model (AM), pronunciation model (PM), and language model (LM) which rely on laboriously engineered processing pipelines. In contrast end-to-end speech recognition systems do not need hand-designed components. Our approach uses well optimized RNNs that uses GPU to train the model.
-
Offline Urdu Image to Text Recognition/Writer’s Gender Identification
Summary: This project has two main features: Offline URDU optical character recognition. Writer’s Identification for urdu text. Offline Urdu character recognition means it will predict the written text on an image. The word “offline” means it will predict the already written text in an image. While “online ocr” means ocr that predicts the character while writing on real time. There are many OCR available for English. But there are very few for urdu and those which are available uses an old approach of “many notations of single character” like below. My project uses End to end solution. It is using deep-learning techniques to predict the characters.
-
** Estimation of the Remaining Useful Life (RUL) of Bearings**
Summary: Having precise estimates of when ball bearings should be replaced can help achieve higher efficiency and ensure safety. We used ball bearing’s vibrations to estimate their remaining useful life (RUL). Features from time, frequency and time-frequency domain were extracted. Their selection was done according to their correlation and monotonicity with time. These were then fused together with LSTM. Double exponential smoothing gave the RUL.
-
Predicting E-Commerce Order Returns - An application of Machine Learning
Summary: One of the biggest challenge that an e-commerce venture is facing on a daily basis is the order return and cancellation rate. These returns not only cost delivery charges but also the opportunity cost of selling product through a valid order. In this paper, we will use historic online purchase data obtained from one of the biggest e-commerce platform in Pakistan and try to predict products that will be returned or cancelled. These predictions will then allow us to intervene in the product delivery cycle well in advance and avoid the costs associated with returned/cancelled orders.
-
Optimal Detection of Side Channel Attacks on Cache Memory using Machine Learning
Summary: In computing, active data is often cached to shorten data access times, reduce latency and improve input/output. Because almost all application workload is dependent upon I/O operations, caching is used to improve application performance. Various attacks like prime+probe, flush+reload and flush+flush are purposed to steal important information by targeting caches using side channel methods. Flush + flush is stealth in nature and not detectable using available methods. I am designing tool to detect flush+flush.