Navid Rajabi

I'm a PhD Candidate in Computer Science (AI/ML) at George Mason University and I have been lucky to be advised by Prof. Jana Kosecka.

My research lies at the intersection of Natural Language Processing (NLP) using Large Language Models (LLMs), Computer Vision, and Robotics/Navigation, aiming for efficient adaptation and development of Multimodal Large Language Models (MLLMs).

More specifically, my research focuses on fine-grained concepts understanding (e.g. grounding and spatial reasoning). Ultimately, my work aims to improve the Embodied AI agents performance for instruction following in photo-realistic 3D environments.

Prior to joining George Mason University, I was a graduate student at the University of San Francisco and Illinois State University , where I received my M.Sc. degree. During my master's degree, I was mainly working on ML and Deep Learning for the Internet of Things (IoT) robustness.

Email  /  LinkedIn /  Google Scholar  /  Github /  Twitter

• Oct 2024: GSR-BENCH: A Benchmark for Grounded Spatial Reasoning Evaluation via Multimodal LLMs has been accepted to NeurIPS 2024 Compositional Learning Workshop.

• Apr 2024: Q-GroundCAM: Quantifying Grounding in Vision Language Models via GradCAM got accepted to WFM @ CVPR 2024.

• Mar 2024: Towards Grounded Visual Spatial Reasoning in Multi-Modal Vision Language Models got accepted to DMLR @ ICLR 2024.

• Nov 2023: Our paper on Bootstrapping Indoor Scenes Annotations with Pseudo-Labels using the Fusion of Vision Foundation Models got accepted to 2024 IEEE/CVF WACV 2nd Workshop on Pretraining.

• Oct 2023: Our paper on Efficient Adaption of LLMs for Machine Translation got accepted to WMT 2023.

• Jun 2022: Joined Qualtrics iQ R&D as ML Applied Scientist Intern for Summer 2022.

• Apr 2022: Passed my PhD Comprehensive Exam on "Vision-and-Language Understanding for Embodied AI".

• Jun 2021: Joined Capital Group as ML Engineer Intern for Summer 2021.

• May 2021: Our paper on Fairness and Equity in Multilingual Question Answering got accepted to ACL 2021.

The objective of my main research project is to enhance the Success Rate Weighted by Path Length (SPL) of the embodied agent in unseen environments by boosting the (1) language grounding and (2) planning/navigation components.

In this project, we measure the effect of employing BERT-based generated embeddings and substituting the LSTM with a Transformer on the Image Captioning task. As a result, we achieved 13.5% higher performance (in terms of BLEU-4) compared to the baseline and noticed Out-of-Distribution object classes/words among the generated captions, due to the usage of BERT.

In this project, we only considered frontal studies and measured the performance of Deep Convolutional Neural Networks (pre-trained on ImageNet) for multi-class (14) binary classification as diagnosis (e.g. Pneumonia, Lung Opacity, Fracture etc).

In addition, we tried to implement the entire pipeline using ML approaches (i.e. without utilizing ConvNets for feature extraction) to improve the interpretability of the model. To be more specific, an unsupervised Stacked Auto Encoder (SAE) was trained on the images and the middle layer was taken as a latent embedding. Then, this embedding was fed into a Random Forest classifier. However, the performance was far behind the Deep Learning approach.

The core competence of this work stems from the fact that this model has been trained, validated and tested on more than 7 million instances of an actual IoT network traffic (consisting of 9 real IoT devices), instead of simulated/synthetic traffic datasets.

We've proposed a more efficient shortcut to interact with the large-screen smart phones with one hand, which outperforms two of current approaches, in both quantitative and qualitative metrics. [NSF Grant]

The main objective of this project is to detect abnormalities at the edge as a security-oriented proactive approach to recognize the future abnormalities in time-series voltage dataset of a given household.

The main idea behind this model is to behave the transactions of a local IoT network as a distributed ledger of Blockchain by employing the Elliptic Curve Digital Signature Algorithm (ECDSA), which is being used in the Bitcoin architecture for transactions integrity.

Selective Exposure is a destructive phenomenon occurs due to the over usage of Recommender Systems. In this work, we've demonstrated the existence of this negative effect using the real, proprietary dataset we've collected from students.

Website design courtesy: Jon Barron