Mahendran Narayanan

I am a Researcher with a working experience of three years. I wear many hats within a short time period. My work primarily focuses on research, and I have also worked as an ML engineer and MLOps test engineer across three projects. Created POC for 2 research projects. Worked across 5 projects till date. I contribute to open source projects in my leisure time.

Areas worked: Network engineering, Vision Transformers, Image classification, machine unlearning, Finetuning LLMs, Generative AI, Regularization techniques (DL), Hyperparameter tuning, Model inference Testing.
Areas open to: Deep learning, LLMs, Generative AI, Computer vision.
News

2024 Paper accepted in conference
17 Nov, 2023 SENetV2: Aggregated dense layer for channelwise and global representations paper in ArXiv (WACV 2024 rejected)
2023 Reviewer for WACV 2024.
26 Aug, 2023 Submitted Handwritten image augmentation paper in ArXiv
Research Google scholar
SENetV2: Aggregated dense layer for channelwise and global representations [ WACV 2024 (Rejected) ]
Proposed a novel aggregated dense layer in a neural network.
Addition of multi dense layers enhance learning representations in SENet.
Variations of Squeeze and Excitation networks [ ArXiv ]
Proposed five variants of Squeeze and excitation module.
This change in module structure impacts the accuracy to a certain extent in SEResnet.
Deep learning for Fitness [ ArXiv ]
Authoring an application paper to verify the body posture in real-time with a reference image and provide immediate feedback and suggestions.
Can be applied effectively in exercise and yoga sessions to ensure participants maintain correct postures, offering real-time feedback and guidance for improved performance and alignment.
Handwritten stroke augmentation on images [ CVPR 2022 (Rejected)]
Authored a novel OCR based universal augmentation technique for OCR texts irrespective of language.
Memory visualization framework for training neural network [ ArXiv ]
Created a web based UI for monitoring memory consumption w.r.t hyperparameters in network.
Identifying tourist destinations from movie scenes using Deep Learning [ ArXiv ]
Developed a methodology along with a curated dataset for identifying travel destinations depicted in movies based on the scenic elements portrayed.
Project Experience

Toxicity Reduction in GPT
Implementing the latest prompt tuning technique to mitigate toxicity in generated content. [Ongoing]
Generate new backgrounds Code
Replacing the background of the generated image.
Leverage Stable Diffusion and Segment Anything models to perform background replacement on the generated image.
GPT unlearning Code
Employing machine unlearning techniques to induce the unlearning of alphabets within GPT.
Curated a dataset focused on alphabets to negatively influence the generated text.
Outcomes denoted A,B,C no longer represent alphabets by the resultant GPT model.
Tamil MNIST dataset
Collected a dataset consisting of more than 36,000 samples for Handwritten OCR specifically tailored for the Tamil language.
Replicating papers
Replicating network engineering papers within the domains of image classification and LLMs.
Experimenting with the latest modules and open-sourcing the project on GitHub.