Projects

Low-Resource Translation as Language Modeling

Published in WMT20

Submission to the very low resource supervised machine translation task at the Fifth Conference on Machine Translation. The goal of this task is to create a system which translates between German and the low resource language Upper Sorbian. Decoder-only transformer architecture is used and the translation task is formulated as language modeling. To address the low-resource aspect of the problem, the model is pretrained over a similar language parallel corpus. Then, an intermediate back-translation step is employed before fine-tuning. Finally, an analysis of the system’s performance is presented.
PDF

Perceptual Image Similarity for Unsupervised Representation Learning
Neural networks have the potential to learn useful features from raw data and successfully use them for a variety of downstream tasks. The performance of a machine learning model relies upon the quality of the learned representations. This work focuses on improving the performance of unsupervised representation learning models by adding image generation as an extra task and using perceptual image similarity as a loss function. Our experiments show that our method works on different neural network architectures and on different datasets. Our method achieves state of the art representation learning performance for AlexNet on CIFAR10, CIFAR-100, SVHN and yields significant improvement on ImageNet.
PDF

Co-Training Images
Co-Training is a machine-learning algorithm that is proposed by Blum and Mitchell [1]. It can be used when a small portion of a dataset is labeled. The original work used the Co-Training algorithm for classifying web-pages. This novel project considers the problem of image classification on CIFAR-10 dataset using Co-Training.
Code  

Multi-party WebRTC Videoconferencing over Software-Defined Networks
Our aim is to propose an architecture and implementation of third-party video service providers (VSP) to offer WebRTC videoconferencing services at a predictable and stable quality level in collaboration with network service providers (NSP) over a multi-operator SDN environment. Our multi-party WebRTC videoconferencing architecture uses a selective forwarding unit (SFU) where all peers send their video streams, and all streams are distributed to peers. In our architecture and implementations, clients perform motion-adaptive layer selection to adapt their send rates to the bandwidths reserved between the endpoints, and clients communicate through Google Chrome browser that supports VP9 codec, which is the most popular codec for scalable video coding and real time video communications. Our experiments show that the proposed framework yields excellent results with stable video quality when compared with the default service in terms of video quality parameters.
PDF   Code

Bluetooth 5.0 Car-to-Car Communication with OBD-II
Since the discovery of radio, wireless communication technologies are tremendously changing the way we live and the way we interact with our environment. Thanks to the discovery of new and better tools like cellular networks, Wi-Fi and Bluetooth; the world today is a more connected place than it has ever been. Interestingly, these approaches to connectivity that are shaping today's personal electronics and household products, did not penetrate well to some industries; like the automobile industry, due to their limitations.
In the wake of the development of these technologies that offer more efficient, superior connectivity, this project aims to explore how we can incorporate emerging connectivity tools; namely the newly released Bluetooth 5.0 standard which offers greater connection range and throughput compared to its predecessors, to establish connection between moving vehicles aiming to provide motorists with a safer and more pleasant ride. It's two main objectives are establishing a channel for verbal communication between fellow motorists and to make sharing on-board diagnostics data; such as current speed, fuel level or engine status between their vehicles possible.
PDF   Code

References
[1] Avrim Blum and Tom Mitchell. 1998. Combining labeled and unlabeled data with co-training. In Proceedings of the Eleventh Annual Conference on Computational Learning Theory, COLT' 98, pages 92-100, New York, NY, USA. ACM. https://www.cs.cmu.edu/~avrim/Papers/cotrain.pdf