الٓمٓ
Tashin Ahmed
তাশীন আহমেদ। ˈtæʃɪn; Name of origin: Tashin (Hausa) - Rising & Ahmed (Arabic) - Highly praised/Magnificent
🔰 AI Researcher & Engineer 〄 with a performance history in AI competitions.
In pursuit of building T shaped skills in AI.
[tashinahmed] under domain [aol.com]
I like to work simultaneously on fundamental R&D in terms of systems (along with data crunching to terminal solution) including Neural Network Architectures (on every domain, i.e. Computer Vision, Natural Language Understanding, Sequential Learning in AI and so on). Furthermore, I want to actively participate in the True AGI revolution since I think my areas of long-term research interests will play a crucial role in this movement.
Previously, I worked as a Data Scientist at AriSaf Tech Japan K.K.. I was a Research Assistant at Dept. of CSE, United International University where I worked on the project, BWriterGuide, funded by ICT Ministry & part of ICT Innovation Fund. I also served as a Research Volunteer at Collaborative Robotics Lab - University of Virginia. Before that, I have completed my internship at CSIR - Central Drug Research Institute, Uttar Pradesh, India.
Industrial Experiences: Classical Computer Vision, Orthophoto Analysis and Segmentation, Quantitative Analysis, Algorithmic Trading, High Frequency Trading, Optical Character Recognition (English, Bengali, Japanese), Medical Image Analysis, Board Game AI.
Long-term Research Interests: Neuro-Dynamic Programming, Cellular Automaton, Swarm (AI) Intelligence, Multimodal & Reincarnating RL, Meta Learning, Core Neural Network Architectures.
Interested in discussing potential options for academic and industrial collaborations.
Experiences
Professional Experiences
For details visit here.
January 2023 - Present (Contractual)
AutoScan: Car damage detection system from Germany. In a controlled environment, detect various sizes of dents from various parts of a car.
- Created a custom data labeling tool tailored for AutoScan's specific needs.
- Translated a previous C++ program to adapt to the new Python DL-based version for car part detection.
- Along with video frames, adapted IR sensor data to develop a rule-based system for detecting car panel areas more effectively than an AI model, resulting in cost savings.
- Prepared an extensible training and testing codebase that works with Nomnio's software, which communicates with the edge device, an arch with 5 distinct camera angles and IR sensors.
- Overcame difficulties in a complex environment by utilizing cutting-edge (SOTA) AI architectures and frameworks for object detection and tracking, as well as upgrading the current SOTA system through core architecture modification.
ARPA - Agriculture & Rural Payments Agency: Orthophoto analysis for agro based landcover detection on whole Malta.
- Processed large GeoTIFF files (~40GB each) for training SOTA AI models for landcover segmentation.
- Focused on agricultural applications, primarily identifying cultivable land, water bodies, and associated features.
- Experimented with cutting-edge frameworks and methods utilzing MS Azure AI platform to address challenges within multi bands dataset chain.
AirImpact: Above ground Biomass detection with AI.
- Above ground Biomass detection with AI. - Conducted tree counting with DL from orthophotos (planetscope data) & prepared a PoC (AI end to end model architecture) for above-ground biomass (AGBM) detection.
hubble_s: Behavioral analysis with AI technology.
- Provided theoretical solution on the AI architecture and support for hubble_s, a company working on building a behavioral operating system (Behavioral OS).
Colleagues and advisors: Dr Leander Grech, Dr Mark Borg and Dr Jean Paul Ebejer.
May 2021 - December 2022 (Full-Time)
TradeX: Automated trader for stock, ForEx and Crypto (Internal Project).
- Worked on the Core Engine of TradeX.
- Developed a rule-based system alongside an AI system that is more profitable than the AI system, which helped the company save a significant sum.
- Brainstormed on in-house product development with superiors.
Junior Software Engineer at AriSaf Tech Japan K.K.
November 2020 - April 2021 (Full-Time)
Say Halal: Halal Packaged item detection and a super app for Muslims in Japan (Internal Project).
- AI-based in-house product development.
- Data Analysis, preparation for ML model (Japanese OCR).
- Rule based string processing model for efficient multi class classification on natural language based dataset that acquired tremendous improvement over ML model.
Colleagues and advisors: Toru Yoshihara and Jahidur Rahman.
February 2020 - June 2020 (Contractual)
Worked on a govt. funded project: BWriterGuide. As part of my work on a number of modules, I was tasked with determining and resolving the existing shortcomings of Bangla/Bengali OCR. The strategy relied on GAN and improved its performance with the help of CycleGAN and other tuning techniques. Deployment of all necessary modules for the project with Heroku and a custom webpage. A part of the work eventually published here with some adjustments.
Supervisor - Dr Chowdhury Mofizur Rahman. Co-Supervisor Chowdhury Rafeed Rahman.
December 2019 - February 2020 (Internship)
Worked with Dr Sukant Khurana with his data analysis team on Multiple Sclerosis disease in the field of Computer Vision for Medical Image Analysis. Worked on the hypothesis of different lesion features specifically, Maximum length of co-ords, Area, perimeter, major and minor axis length, eccentricity, convex area, volume & how their changes can affect EDSS scores.
Supervisor - Dr Sukant Khurana.
Volunteer Experiences
Till July 2024
November 2020 - 2022
Mostly I lent my hands on bug hunting for neurips.cc website till now. Tested environments like gather town for virtual conferences.
June 2020 - October 2020 (Apprenticeship)
Worked with Md Mofijul Islam on a multimodal based action recognition learning model. Competed in Hateful Memes competition hosted by FB & DrivenData.
Content Tester & Community Leader at DeepLearning.AI
March 2022 - November 2022
Tested resources (video lectures, codes, manuals) for unreleased courses.
KaggleX BIPOC Mentor (Cohort 3) Peer Reviews
Publications
Published
Pipeline Enabling Zero-shot Classification for Bangla Handwritten Grapheme
Linsheng Guo, Md Habibur Rahman Sifat, Tashin Ahmed (Initiator)
Accepted for - EMNLP 2023 - Workshop
on Bangla
Language Processing (BLP) (Oral Presentataion, Track: Main - Long)
Evaluating The Effectiveness of Capsule Neural Network in Toxic Comment Classification using
Pre-trained BERT Embeddings
Md Habibur Rahman Sifat, Noor Hossain Nuri Sabab, Tashin Ahmed (Initiator and
directional lead)
Accepted for - TENCON 2023 - IEEE Region
10 Technical
Conference (Main Conference : Oral Presentation)
Dual phase convolutional neural network based system aimed at small rice grain dataset for disease
identification
Tashin Ahmed (bachelor’s thesis, lead), Chowdhury Rafeed Rahman, Md. Faysal Mahmud Abid
Accepted for - AAAI 2023 - AI for
Agriculture and Food Systems (Workshop : Lightning Talk + Poster)
Referred at -
Geographical Research Bulletin
Automatic Signboard Detection and Localization in Densely Populated Developing Cities
Md Sadrul Islam Toaha, Sakib Bin Asad, Chowdhury Rafeed Rahman, SM Haque, Mahfuz Ara Proma, Md Shuvo, Ahsan
Habib, Tashin Ahmed (co-supervisor), Md Basher
Published in - Signal Processing: Image Communication
Classification and understanding of cloud structures via satellite images with
EfficientUNet
Tashin Ahmed, Noor Hossain Nuri Sabab
Published in - Springer
Nature Computer Science (SNCS)
Preprints + Ongoing Works + On Hold
- Teaching Convolutional Neural Networks to Solve Conway’s Reverse Game of Life
- Bengali Lip Reading from Video Dataset with 3D Convolutional Neural Network
- Automated Market Trend Forecasting and Order Placement: A Hybrid Approach integrating Multiple AI Systems
- Collective Wisdom in Language Models: Unveiling the Power of LLM-Swarm
- Exploring Adaptive Memory Patterns in Reinforcement Learning Agents
Competitions
Best Performances
Basically, I tried to create memory patterns for different role set i.e. defence, offence, and attack. Different RL methods have been tested for different roles. Highest skill point gained in leaderboard μ=1169.9.
Tried to put different categories of bots that are correlated to Swarm Intelligence by utilizing different NN techniques. Highest gain on the LB was μ=1085.5.
Competition was about beaviorial representation learning from ants and beetles using video dataset. We have worked with PyTorch SimCLR a Simple framework for Contrastive Learning of visual Representations for our model.
Projects
Research Projects
AutoScan AI Scanner
SOTA Object Detection and Tracking system to detect multiclass objects in controlled environment. Classical Computer Vision approach taken in terms of AI failure that saves cost both in terms of money and time.
ARPA-Orthoseg
Othophoto Analysis on bigdata for multiclass classification and segmentation for agro based project.
Project for Agriculutre and Rural Payment Agency of Malta.
Dhan
Oushodhi
An AI-based mobile phone application (Android) for detecting different Rice Grain diseases in context of Bangladesh.
Undergraduate Capstone Project.
BWriter Guide
A user-friendly Bangla writer web application which will have many options for the user to choose from.
This project is funded by ICT Ministry, Bangladesh & part of ICT Innovation Fund.
Say Halal
AST's internal research and development project for a mobile phone app that uses Japanese OCR and a natural language-based classification system.
Games & Simulations (Academic Projects)
TETRIS
Single player Tetris game created in C++ & Java both. Graphics added for both version.
10
different
levels
are created depending on play time with addictive scoring system.
LIFE
Zero-player
game
for
observing cellular
automaton & evaluation of their
growth. Recreation of Conway's game of
Life using C language & OpenGL.
MAZE
A maze creator algorithm which creates random mazes using OLC Engine & have GUI
written in C++ comes with automated maze solver plus a game to solve maze puzzle. Both
program
tested on
TURBO C++.
Certifications
MOOCs
DeepLearning.AI Tensorflow Developer: Professional Certificate
deeplearning.ai
Grade: 93.785%
Data-driven Astronomy
University of Sydney
Grade: 89%
Social Psychology (with Honors)
Wesleyan University
Grade: 92%
Others
Outstanding Reviewer
First Workshop on Bangla
Language Processing (BLP) - EMNLP 2023
Certification
(Private)
Duolingo English Test (w/o preparation scored 140/160, IELTS comparative score -
7.5/9)
Duolingo English Test
Certification
(Private)
Triplebyte Certified
Triplebyte
Certification
Eterna (Solve Puzzles, Invent Medicines)
Achievements
Certification
Others
Write Ups
Time Series
Analysis
(TSA) for beginners
You have access to 10 notebooks (created using Kaggle Notebook) to get you started with
Time
Series
Analysis (TSA). Continue reading
to find out more.
TL;DR
All ten of the notebooks are briefly described in this discussion, along with links to
each
notebook.
A variety of arbitrary pattern generation strategies, validations, label construction,
and
diverse
ML and DL techniques, including manually
constructed CNN, DNN, LSTM, and RNN using TensorFlow Sequential layers—all of which are
ideal
for
beginners—were tried to be provided.
The last two notebooks demonstrate basic EDA and the application of DNN and CNN to the
Sunspots
dataset.
Sunspots
Dataset: Sunspots are transient occurrences on the photosphere of the
Sun that
appear as spots that are darker than their surroundings. These are areas of lower
surface
temperature brought on by magnetic flux concentrations that prevent convection. Sunspots
typically
come in pairs with completely different magnetic polarity. (incomplete)
Introduction to Genetic Algorithms (GA)
This blog post is an assembly of GA from the perspective of biological anecdote to a
deep dive
on it's history, simple architecture and use cases.
Capsule Neural Network (CapsNet): A Forgotten ANN
Architecture
CapsNets are artificial neural networks (ANN) that better capture
hierarchical relationships. They’re inspired by biological neural arrangements. CapsNet
use “capsules”, which reuse outputs
from multiple capsules to stabilize representations for...
