Publications

Peer-reviewed papers and preprints, most recent first. A complete and continuously updated list is available on Google Scholar.

2026

FS-DFM: Fast and Accurate Long Text Generation with Few-Step Diffusion Language Models

Amin Karimi Monsefi, Nikhil Bhendawade, Manuel Rafael Ciosici, Dominic Culver, Yizhe Zhang, Irina Belousova

ICLR 2026 — The Fourteenth International Conference on Learning Representations, Rio de Janeiro, Brazil, 2026

Diffusion language models promise parallel, controllable text generation but usually need many refinement steps to match autoregressive quality. FS-DFM is a few-step discrete flow-matching model that makes the number of sampling steps an explicit parameter, so long-form text can be generated accurately with a small, fixed step budget instead of hundreds of iterations.

Paper

2025

ISOSNet: A Unified Framework for Cone Photoreceptor Detection and Inner Segment and Outer Segment Length Measurement from AO-OCT B-Scans

Mengxi Zhou, Yue Zhang, Eli Kirkendall, Amin Karimi Monsefi, Matthew Wolfe, Kiran A. Chitkara, Stacey S. Choi, Nathan Doble, Srinivasan Parthasarathy, Rajiv Ramnath

Biomedical Optics Express (Optica Publishing Group), 2025

ISOSNet is a unified deep-learning framework that detects cone photoreceptors in adaptive-optics OCT B-scans and measures inner-segment and outer-segment lengths in a single pass, replacing a slow and subjective manual annotation workflow with an automated, reproducible one.

Paper

TaxaDiffusion: Progressively Trained Diffusion Model for Fine-Grained Species Generation

Amin Karimi Monsefi, Mridul Khurana, Rajiv Ramnath, Anuj Karpatne, Wei-Lun (Harry) Chao, Cheng Zhang

ICCV 2025 — International Conference on Computer Vision, Honolulu, Hawai'i, USA, 2025

TaxaDiffusion trains a diffusion model along the biological taxonomy — from coarse ranks such as class and order down to species — so that the model first learns shared morphology and then the subtle traits that separate visually similar species. The progressive, taxonomy-aware curriculum improves fine-grained generation and makes the learned trait hierarchy inspectable.

Paper

KnobGen: Controlling the Sophistication of Artwork in Sketch-Based Diffusion Models

Amin Karimi Monsefi, Pouyan Boreshnavard, Mengxi Zhou, Wei-Lun (Harry) Chao, Alper Yilmaz, Rajiv Ramnath

CVEU Workshop, CVPR 2025, Nashville, Tennessee, USA, 2025

Sketch-based diffusion models tend to assume a skilled artist: they follow detailed sketches well but struggle with rough, novice drawings. KnobGen adds a single control — a “knob” — that adjusts how strictly generation follows the sketch, so the same model serves both quick doodles and precise, expert line art.

Paper

DetailCLIP: Detail-Oriented CLIP for Fine-Grained Tasks

Amin Karimi Monsefi, Kishore Prakash Sailaja, Ali Alilooee, Ser-Nam Lim, Rajiv Ramnath

SSI-FM Workshop, ICLR 2025, Singapore, 2025

Contrastive vision-language models such as CLIP learn image-level semantics and lose the pixel-level detail that segmentation and other dense tasks depend on. DetailCLIP combines patch-level self-distillation with a reconstruction objective so that a CLIP-style encoder keeps fine spatial detail while retaining its semantic strength.

Paper

Frequency-Guided Masking for Enhanced Vision Self-Supervised Learning

Amin Karimi Monsefi, Mengxi Zhou, Nastaran Karimi Monsefi, Ser-Nam Lim, Wei-Lun (Harry) Chao, Rajiv Ramnath

ICLR 2025 — The Thirteenth International Conference on Learning Representations, Singapore, 2025

Masked image modeling usually masks patches at random, which ignores where the informative signal actually lives. This work uses the frequency content of an image to guide masking, producing a pre-training objective that learns stronger representations with markedly less pre-training data and compute.

Paper

2024

Reducing Manual Labeling Requirements and Improved Retinal Ganglion Cell Identification in 3D AO-OCT Volumes Using Semi-Supervised Learning

Mengxi Zhou, Yue Zhang, Amin Karimi Monsefi, Stacey S. Choi, Nathan Doble, Srinivasan Parthasarathy, Rajiv Ramnath

Biomedical Optics Express (Optica Publishing Group), 2024

Identifying retinal ganglion cells in 3D adaptive-optics OCT volumes normally demands extensive expert annotation. This work uses semi-supervised learning to exploit unlabeled volumes, cutting the manual labeling burden while improving identification accuracy.

Paper

Masked LoGoNet: Fast and Accurate 3D Image Analysis for Medical Domain

Amin Karimi Monsefi, Payam Karisani, Mengxi Zhou, Stacey Choi, Nathan Doble, Heng Ji, Srinivasan Parthasarathy, Rajiv Ramnath

KDD 2024 — 30th ACM SIGKDD Conference on Knowledge Discovery and Data Mining, Barcelona, Spain, 2024

In this paper, we introduce a new neural network architecture, termed LoGoNet, with a tailored self-supervised learning (SSL) method to mitigate such challenges. LoGoNet integrates a novel feature extractor within a U-shaped architecture, leveraging Large Kernel Attention (LKA) and a dual encoding strategy to capture both long-range and short-range feature dependencies adeptly. This combination of strategies is especially beneficial in medical image segmentation, given the difficulty of learning intricate and often irregular body organ shapes. Complementarily, we propose an SSL method tailored for 3D images to compensate for the lack of large labeled datasets.

Paper PDF

2023

CrashFormer: A Multimodal Architecture to Predict the Risk of Crash

Amin Karimi Monsefi, Pouya Shiri, Ahmad Mohammadshirazi, Nastaran Karimi Monsefi, Ron Davies, Sobhan Moosavi, Rajiv Ramnath

UrbanAI '23 — 1st ACM SIGSPATIAL International Workshop on Advances in Urban-AI, Hamburg, Germany, 2023

We propose CrashFormer, a multi-modal architecture that utilizes comprehensive (but relatively easy to obtain) inputs such as the history of accidents, weather information, map images, and demographic information. The model predicts the future risk of accidents on a reasonably acceptable cadence (i.e., every six hours) for a geographical location of 5.161 square kilometers.

Paper PDF

Novel Physics-Based Machine-Learning Models for Indoor Air Quality Approximations

Ahmad Mohammadshirazi, Aida Nadafian, Amin Karimi Monsefi, Mohammad H. Rafiei, Rajiv Ramnath

KDD 2023 — 29th ACM SIGKDD Conference on Knowledge Discovery and Data Mining, Long Beach, California, USA, 2023

The study introduces six novel physics-based machine learning models designed to accurately estimate indoor pollutant concentrations using cost-effective sensors, leveraging domain knowledge through a combination of physics concepts, Gated Recurrent Units, and Decomposition techniques, showcasing their superiority in terms of computational efficiency and accuracy over transformer-based models using real-world office data.

Paper PDF

Smart And Collaborative Industrial IoT: A Federated Learning And Data Space Approach

Bahar Farahani, Amin Karimi Monsefi

Digital Communications and Networks (Elsevier), 2023

The challenges faced in adopting Industry 4.0, particularly in the field of smart factories and production, due to issues such as lack of high-quality and diverse data, fragmented data across different silos, and concerns regarding privacy and security. To address these challenges, the article proposes a decentralized architecture utilizing multi-party technologies, privacy-enhancing techniques, and AI approaches to create a collaborative platform and federated data space. Experimental results demonstrate the potential benefits of this approach for multi-party applications and data sharing based on the FAIR principles.

Paper PDF

2022

Will There Be A Construction? Predicting Road Constructions Based On Heterogeneous Spatiotemporal Data

Amin Karimi Monsefi, Sobhan Moosavi, Rajiv Ramnath

ACM SIGSPATIAL 2022 — 30th International Conference on Advances in Geographic Information Systems, Seattle, Washington, USA, 2022

A computational approach for predicting future road construction projects by integrating and analyzing various types of spatiotemporal data. The approach utilizes a deep-neural-network-based model trained on a large dataset called “US-Constructions,” which includes 6.2 million road constructions with diverse attributes and road-network features. Experimental results demonstrate the effectiveness of the approach in accurately predicting future road constructions in several major cities in the United States.

Paper PDF

2020

Deep Learning Prediction of Heat Propagation on 2-D Domain via Numerical Solution

Behzad Zakeri, Amin Karimi Monsefi, Babak Darafarin

The 7th International Conference on Contemporary Issues in Data Science, 2020

In This work demonstrates the effectiveness of deep learning in solving the problem of two-dimensional heat transfer in an arbitrary domain. Using the finite volume method, the researchers trained a deep neural network on 100,000 cases to predict the heat transfer solution. The results show that the network achieved satisfactory precision compared to the commercial program ANSYS, indicating the potential of deep learning in accurately predicting physics-based heat transfer phenomena.

Paper PDF

2019

Performing Software Test Oracle Based on Deep Neural Network With Fuzzy Inference System

High-Performance Computing and Big Data Analysis: Second International Congress, TopHPC 2019, 2019

This work introduces a novel software test Oracle based on deep learning and a fuzzy inference system, aimed at automating the testing process while minimizing time and cost. The Oracle maps the software output to a fuzzy space using Takagi-Sugeno-Kang fuzzy inference and trains a deep neural network. The performance of the Oracle is evaluated using different models, demonstrating its ability to accurately detect correct and false results.

Paper PDF

Solving The Reaction-Diffusion Equation Based On Analytical Methods And Deep Learning Algorithm; The Case Study Of Sulfate Attack To Concrete

Amin Karimi Monsefi, Rana Bakhtiyarzade

arXiv preprint, 2019

In this study, a deep neural network was trained to predict the solution of the reaction-diffusion equation with varying coefficients, utilizing numerical and analytical solutions. Dimensional analysis technique was employed to reduce learning time and identify similar equation solutions. The results show that deep learning successfully estimated the solution of the reaction-diffusion equation with a constant coefficient, highlighting its accuracy in solving partial differential equations.

Paper PDF

Real-time Travel Time Estimation Using Matrix Factorization

E. Badrestani, B. Bahrak, A. Elahi, A. Faramarzi, P. Golshanrad, A. Karimi Monsefi, H. Mahini, A. Zirak

arXiv preprint, 2019

This paper addresses the estimation of travel times for various road segments and time intervals using GPS data and Matrix Factorization techniques. By aggregating GPS data into a matrix and applying the Alternating Least Squares (ALS) method with regularization, the proposed approach effectively solves the sparsity problem and accurately estimates travel times. Evaluation results using real data from a large online taxi dispatching system in Iran demonstrate the strength and effectiveness of the proposed method.

Paper PDF

Weakly Supervised Learning Technique for Solving Partial Differential Equations; Case Study of 1-D Reaction-Diffusion Equation

High-Performance Computing and Big Data Analysis: Second International Congress, TopHPC 2019, 2019

This paper introduces a new method that utilizes weak supervision and deep learning to solve partial differential equations (PDEs) using only boundary and initial conditions, making it suitable for unknown PDEs without labeled data. The approach is evaluated by solving the Reaction-Diffusion equation, demonstrating high consistency with the finite difference method and highlighting the effectiveness of weakly supervised learning in solving various types of differential equations.

Paper PDF