-
Learning Robust Latent Representation in Speech
We make architectural modification for a Transformer based VAE. The proposed VAE is not only robust against latent cluster collapse in case of limited data, but also discovers rich features for controllable speech synthesis in Tacotron-2 based speech generative models.
[ ACL2021 Findings ] -
Adaptive Transformers in RL
In this work we showed that when a learnable context length is used for attention in partially observable and memory intensive tasks, it surpasses the performance of the current state-of-the-art Transformer XL and XL-1 architectures. To our knowledge this was the first work to stabilize Adaptive Attention Span in RL.
[ arxiv preprint ] [ Codes ] -
Adaptive Attention based Kernels for Image Classification
In this paper we propose an algorithm for learning self-attention kernel sizes in computer vision and compare its performance to fixed-size local attention and convolution kernels. We then discuss whether adaptive attention can be helpful in correlating global features and lead to any reduced FLOP count over CNN and attention based architectures.
[ arxiv preprint ] [ Codes ] -
Approximate Cross Validation in Incremental Learning
During my final year of Bachelors, I worked with Dr. Shobha G., Sneha M. and Rahul MV in inventing method for functional estimation of cross validation in incremental learning models. This research eliminated the need to store previous data points in online learning systems for cross validation. -
Contention Control in Machine-to-Machine Communication
In 2016, I worked with Dr. BR Tamma, Dr. MK Giluka and Dr. AA Franklin to propose an extreme Random Access Channel congestion control algorithm for machine to machine communication with LTE-A as the backbone network. Our proposed algorithm had 20% greater success probability than the then 3GPP state of the art.
[ 2018 IEEE 4th World Forum on Internet of Things ]
Teaching
-
CSC384 Introduction to Artificial Intelligence (Winter 2020)
Designing tests and assignments on probability, bayesian networks, markov models and monte carlo tree search methods. The course is being supervised by Prof. Bahar Aameri and Prof. Sonya Allin.
[ Coursepage ] -
CSC108 Introduction to Python Programming (Fall 2019)
1st year introductory course on python programming for all CS and non CS students. The course was supervised by Prof. Jennifer Campbell and Prof. Mario Badr.