Arjon Das

Email   /   Github   /   Twitter

I am a freelance programmer and Deep-learning practitioner. I completed my bachelor's in Computer Science and Engineering at the Chittagong University of Engineering and Technology. There I worked on IoT Based Smart Food Monitoring System under the supervision of Dr. M. Moshiul Hoque.

I am pursuing my MS in Computer Science with AI Concentration at the University of Nebraska at Omaha. I'm a Graduate Research Assistant at UNO's RNA Lab under the supervision of Dr. Xin Zhong. I do Computer Vision research, specializing on Self-Supervised Representation Learning.


Self-Supervised Representation Learning

Title Hidden: Work Under Review   

Proposed a new augmentation. Applying with different Joint Embedding Architectures, results in higher linear evaluation accuracy along with Distortion Robustness on previously unseen augmentations. Below chart shows the comparison of VICReg Baseline and VICReg with our proposed augmentation.

Paper Coming Soon

Deep Learning-based Audio in Image Watermarking (Deep AIWM)   

Deep AIWM is a Deep Learning model that can embed audio data inside RGB image without degrading the quality and extract the audio data from the marked image. The model's Siamese Network is able to distinguish noisy extracted audio from the marked image which can be used for speech command recognition simply by scanning image.

Code Coming Soon

Simple Sentiment   PyTorch  |  Flask

Simple Sentiment is a DistilBERT based classifier that can recognize positive and negative vibes from sentences. It was trained using the Sentiment140 Dataset comprising of 1.6 Million tweets labeled positive or negative.

Project Website   /   Fine tuned Model   /   Code

Simple MANN (using LSTM)   TensorFlow

Simple MANN is a meta-learning implementation inspired by the paper Optimization as a Model for Few-Shot Learning. Even though the name suggests MANN(Memory Augmented Neural Networks), instead of NTM this implementation uses a single layer LSTM as memory unit. The model is evaluated on the Omniglot Dataset.

Details + Code

Facebook Hateful Meme Challenge   PyTorch   |  MMF

Submission for Facebook Hateful Meme Challenge at DrivenData using VisualBERT. The model scored 0.7214(AUC-ROC) without ensemble and task specific pre-training. Model's MMF configuration is given below.


COVID-19 Identification from Chest X-Ray   PyTorch

Classifier based on ResNet-18 model which has been finetuned on the Chest X-Rays from COVID-19 Radiography Database. The classifier can differentiate X-Rays of Normal Condition, Viral Pneumonia and COVID-19 patients.

Details + Code

Stock Scraper   MongoDB  |  Express  |  React.js  |  Node.js

Stock Scraper is a tweet scraping server that counts number of times any stock has been tweeted by targeted users at Stocktwits and visualizes the data. This project is a excellent groundwork for analyzing and extracting values from social investment platforms.

Project Website

IoT Based Smart Food Monitoring System   
Raspberry Pi  |  Swift  |  Python  |  MongoDB  |  Express  |  Node.js

The project facilitates food monitoring of refrigerators and food cold storage systems using mobile devices, with high scalability. The system is developed using Raspberry Pi device and transformed into a small form factor IoT device. One of the pros is it's ability to scale up which can facilitate large scale cold storage monitoring.

Thesis   /   Code

Code reproduction of various papers   PyTorch

Code reproducibility of papers is becoming more and more challenging as models are becoming more complex and abstract. Here are some of my practice attempts to reproduce models from papers with help from various online forums using popular machine learning frameworks.


Online Job Market (Job Engine)   React.js  |  Javascript

Job Engine is a single page web app for searching and posting jobs online. The browser based app’s front end is implemented using Javascript, React.js and server intergration using REST API.


Car Rental System (EasyRide)   Javascript  |  PHP  |  HTML

Easyride is a database project for renting cars in a particular area, facilitating CRUD functions to the end-user. The project is implemented using PHP, HTML, CSS, MySQL, Javascript.

Details + Code

Direct 2D   Swift

Direct 2D is a self initiated game development project. The game was implemented using Swift with basic physics written from scratch. The app was available on Apple AppStore from July 2017 to July 2018 with over 2.6K downloads.