I spent my summer working at Adobe Research Lab guided by my mentors Sumit Shekhar, Akhash Amarnath and Inderjeet Nair . We worked in the domain of multimodal content for Interactive Documents mainly developing an end to pipeline for revisualization of infographics using state of the art deep learning techniques. My work was more development-based. We performed literature reviews to come up with a problem statement. Due to the lack of structured infographic content, we developed our own custom dataset for training the GAN models. We tried to understand the reading order and revisualized the infographics in formats better suited for mobile devices like the Liquid Mode in Adobe Acrobat Reader. We are looking to patent our project.

Here I was responsible for collecting the stock data of multiple Indian Companies for price prediction based on time series models. After cleaning the data and deriving insights into the kind of data I would be working on, I used the best model adjusting its parameters (for example, AIC) accordingly. I used ARIMA, SES, SARIMAX and Facebook's Prophet to derive good prediction results.

I worked under the guidance of Prof. Siva Theja Maguluri and Prakirt Jhunjhunwala in the topic of Reinforcement Learning-based algorithms for input queued switches in a high traffic setting. I achieved a 5% improvement in the average queue length compared to the optimal MaxWeight approach. I also implemented a Double Deep Q-Learning Network based on the modified algorithm. While I could only run it for a small number of iterations, future work could include training the neural network over many iterations. However, we are yet to understand the dependence of step sizes on the same. CODE REPORT

I worked with Prof. Arkarpava Basu and Shweta Pandey on analyzing workload performance for heterogenous compute and memory systems. The work consisted of using Persistent Memory in GPUs or GPMEM. We worked on building a more robust system to achieve faster throughputs for various workloads. My work was understanding the effects of various workloads on the CPU side of the system and benchmarking the results accordingly.

I worked with Prof. R Manivasakan and Banshree Sarma of the Electrical Engineering Department of IITM on devising a method to implement Deep Reinforcement Learning (DRL) based Resource Allocation in Integrated Access and Backhauled network using Multi-Agent Reinforcement Learning. After creating an environment for Multi-Agent RL simulations, we used Double Deep Q Learning networks and various optimizations to create a model that optimally allocates subchannels and powers to maximize the downlink data rate. The paper was published at the European Conference for Communications Systems (ECCS) 2023. Here is the Arxiv Link to the paper.