Ashim Dahal

I'm an Undergraduate Research Assistant at The University of Southern Mississippi in Mississippi, USA where I work on the Cyber Innovations Lab mostly focusing on modern deep learning techniques on computer vision and multimodal systems.

At the lab, I am advised by Dr. Nick Rahimi and closely work with him and his graduate and undergraduate students on topics like Stable Diffusion, Vision Transformers, Kolmogorov Arnold Networks, Cybersecurity, Visual Question Answering, NLP and Brain-Computer Interface.

I am also the Head of Artificial Intelligence at Google Developers Student Club at USM.

Non-Tech Trivia: I play bansuri and read literature. I also won the Eagles Write Award (best assignment among all freshmen) 2023 for an essay I wrote in 45 minutes.

Email  /  CV  /  Bio  /  Scholar  /  Github /  Linked In

Research

I'm interested in computer vision, deep learning, generative AI, and image processing. Most of my research is about understanding scene, analysing Deep Learning techniques, and inferring meaningful information by tinkering with model architectures and their loss function optimization. Some of my works in computer vision are highlighted.

Threats over X are given on multiple languages. This paper proposes a new dataset and methodology to detect cyber threats spread over tweets on X.

This paper analyzes Convolutional Kolmogorov Arnold Networks on ImageNet with Alexnet, MNIST with LeNet and Tabular CNN modification with MoA datasets.

Analysis of Vision Transformers (ViT) against Convolutional Neural Networks (UNet CNN) for image segmentation on Remote Sensing iSAID dataset. We propose a novel loss function that helps a smaller CNN model to perform equally to a 5x larger ViT model.

Analysing zero day cyber attacks with MLP and SHAP. We use weighted loss function that changes the focus of the model into different parameters while learning thus leading to a model that doesn't overfit on majority class.

We created "Jelly" the first Romanized Nepali Chatbot using facebook's BlenderBot and conducted a survey analysing the efficacy of natural language for mental health conversational bots.

We fine tuned a LeNet style CNN architecture to do OCR on handwritten Devanagari characters. These characters tend to be more complicated to understand than roman numerals and alphabets.

Used TensorFlow's guide on a self curated chicken recipe dataset to then create a GRU based RNN model which would generate new recipes. Did ablation study on data preprocessing and noted down the effects of different processing techniques on the final generator.

Surveyed 300+ individuals regarding the various implications of living on a robot-centric economy. Questions cover topics including privacy, jobs, requirements and feature required for robots to be mainstream or house-help machines.

Miscellaneous

This website is adapted from: Jon Barron.