About Me

Hi! I am Mehrab. I received my PhD in Computer Science specializing in Artificial Intelligence and Machine Learning from the University of California San Diego (UCSD). My advisor was Professor Garrison W. Cottrell. During my PhD, I focused on deep generative models, particularly on debiasing them in various tasks such as attributes to image generation, image-to-image translation, text-to-image generation, etc. Additionally, my research interests include recommender systems, federated learning, and scalable machine learning. To learn more about my projects and relevant publications, please visit my projects and publications, respectively.

Before joining UCSD, I was a research assistant working with Dr. Muhammad Abdullah Adnan at my undergraduate university, Bangladesh University of Engineering and Technology (BUET), where I received my bachelor’s in CS. There I spent a wonderful time developing scalable machine learning algorithms (e.g. PCA) in the distributed environment (e.g Spark/Hadoop). After joining UCSD, I had the opportunity to work as a Data Science intern at Etsy (Summer 2019), Adobe (Summer 2020, 2021, and 2022), and The Home Depot (Fall 2022). I am grateful to Adobe and The Home Depot for the generous gift funds to my advisor for supporting my research.

Latest News

[April 2023] Joined Adobe as a Research Scientist!

[March 2023] Successfully defended my doctoral disseration!

[September 2022] Joined The Home Depot Online Data Science team as a Data Science Intern (part-time, co-op program).

[September 2022] Joint work with Adobe “Discovering and Mitigating Biases in CLIP-based Text-to-Image Generation” accepted as a post at Responsible Computer Vision at ECCV’22.

[August 2022] Joint work with Adobe “Debiasing Image-to-Image Translation Models” accepted in BMVC’22.

[June 2022] Joined Adobe REAL team as a Research Scientist Intern.


Debiasing Generative Models

Discovering and Mitigating Biases in CLIP-based Text-to-Image Generation
Discovered the queries for which the popular CLIP model biases the generated images in the text-to-image synthesis task and proposed several ways to mitigate the biases without retraining CLIP or the underlying generative model.
Tech Stack: Pytorch

Debiasing Image-to-Image Translation Models
Pretrained StyleGAN2 based networks show various biases in different image-to-image translation tasks (such as super-resolution, sketch-to-image, etc.). Mitigated this bias issue using contrastive learning and uniform sampling of minority attributes.
Tech Stack: Pytorch

Bias Detection in Image Search and Mitigation
Identified the bias issue in the image results for search queries, proposed a way to audit. In addition, proposed an attribute-controlled style-based generator to create new content to mitigate such biases and enrich user experience.
Tech Stack: Pytorch, Tensorflow

Recommender Systems

Dynamic Convolution
Built an adaptive convolution network which changes its kernel dynamically depending on the current input (~10% better recommendations).
Tech Stack: Pytorch

Intent Detection for Recommendation
Captured users’ hidden intents (i.e. explore, purchase) from their interactions by designing a hierarchical Transformer model. It first discovers these intents and then pays attention to them for next item prediction (improved personalized recommendations by 5%).
Tech Stack: Tensorflow

Lightweight Convolutional Network for Recommendation
Improved the scalability of sequential recommender methods by modelling a scalable depth-wise separable 1D convolution neural network (requires ~30% less memory).
Tech Stack: Tensorflow

Multi-modal Learning

Visual Commonsense Reasoning.
Enforced reasoning for ans. prediction on VCR by building a differentiable module which jointly trains ans. and rationale prediction (performed better in leaderboard).
Tech Stack: Pytorch

Rationale Generation
Tasked state-of-the-art Visual Question Answering model (ViLBERT) with rationale generation (using GPT-2) to interpret/justify answer prediction. It improves accuracy by 1.5% as well.
Tech Stack: Pytorch

Scalable Machine Learning

Scalable Video Fingerprinting
Built a scalable, end-to-end pipeline using FAISS library that can trace a manipulated video in less than a second from a trusted database with millions of corpuses.
Tech Stack: Tensorflow

Distributed Algorithm Design
Extended both Spark and Hadoop for creating geo-distributed clusters in AWS and designed geo-distributed algorithms for higher dimension data.
Tech Stack: Java, Spark, Hadoop

Scalable Principal Component Analysis.
Improved the scalability of PCA for large datasets (up to 83× better performance) using sketching technique.
Tech Stack: Java, Scala, Spark


Fair Image Search Engine using CLIP
Built a search engine by calculating the cosine similarity scores between CLIP embedding of the text query and images in the database and sorting from highest to lowest to show the most relevant images on the top. Additionally, employed a diversity constraint using the embedding for different demographics.
Tech Stack: Pytorch

Image Colorization using Cycle Consistency Loss
Explored the potential of using Cycle Consistency Loss between grey and colored images in Generative Adversarial Networks for generating true and vivid colors for black & white images.
Tech Stack: Pytorch

Pragmatic Probabilistic Principal Component Analysis
Using sketching techniques, derived a warm initialization for Expectation Maximization (EM) algorithm of Probabilistic PCA (PPCA). This speeds up convergence up to 2.25×.
Tech Stack: Java, Spark

Digitor: A Digital Circuit Simulator
Developed a digital circuit simulator app where one can draw digital circuits and simulate its behavior. It can automatically derive the boolean expression from the circuit and minimize it (using Quine–McCluskey algorithm) to suggest a simplified implementation.
Tech Stack: Java


Md Mehrab Tanjim, Ritwik Sinha, Krishna Kumar Singh, Sridhar Mahadevan, David Arbour, Moumita Sinha
United States Patent Application 17/485780, 2023


For latest publications, please visit my Google Scholar profile.

Discovering and Mitigating Biases in CLIP-based Text-to-Image Generation.
Md Mehrab Tanjim, Krishna Kumar Singh, Kushal Kafle, Ritwik Sinha, Garrison W. Cottrell
Responsible Computer Vision at ECCV (RCV@ECCV), 2022

Debiasing Image-to-Image Translation Models.
Md Mehrab Tanjim, Krishna Kumar Singh, Kushal Kafle, Ritwik Sinha, Garrison W. Cottrell
British Machine Vision Conference (BMVC), 2022

Generating and Controlling Diversity in Image Search.
Md Mehrab Tanjim, Ritwik Sinha, Krishna Kumar Singh, Sridhar Mahadevan, David Arbour, Moumita Sinha, Garrison W. Cottrell
Winter Conference on Applications of Computer Vision (WACV), 2022

Fast, scalable and geo-distributed PCA for big data analytics
TM Tariq Adnan, Md Mehrab Tanjim, Mummad Abdullah Adnan
Information Systems, Elsevier, 2021
pdf | code

DynamicRec: A Dynamic Convolutional Network for Next Item Recommendation
Md Mehrab Tanjim, Hammad A. Ayyubi, Garrison W. Cottrell
Conference on Information and Knowledge Management (CIKM), 2020
pdf | code

Attentive sequential models of latent intent for next item recommendation
Md Mehrab Tanjim, Congzhe Su, Ethan Benjamin, Dian Hu, Liangjie Hong, Julian McAuley
World Wide Web (WWW), 2020
pdf | code

sSketch: A scalable sketching technique for PCA in the cloud.
Md Mehrab Tanjim, Muhammad Abdullah Adnan
Web Search & Data Mining (WSDM), 2018
pdf | code