Current Seed Grants

The key to effective preservation of biodiversity is its estimation, and the ability to follow, and ultimately predict biodiversity changes in space and time. However, the current approaches to biodiversity estimation are slow and costly. They necessitate the classification of observed individuals and the creation of an inventory of taxa. Using automatic classification by deep-learning models is a major step forward, but it still requires manual annotation of large training sets, and the resulted models are limited in scope to the specific fauna and flora of a given location. Due to the high price tag of biodiversity estimation, the data available for decision makers is very sparse. Thus, cost-effective approaches to biodiversity estimation are essential. The current proposal aims to offer one, specifically, focusing on flying insects. Identifying and classifying insects by human experts is difficult and time-consuming, due to their huge number of species and small body sizes. These challenges motivate our search for automated biodiversity monitoring of insects. We propose a collaborative effort to the development of a novel approach to biodiversity estimation. The proposal focuses on estimating the diversity of flying insects, combining a cost-effective, low-tech sampling procedure (sticky traps) and cutting-edge analysis using supervised and unsupervised image processing and machine learning.

Research on the synthesis of geometric meta-structures that are lightweight, strong, and adhere to other performance constraints is a highly active research topic in materials science, geometry, and machine learning. Additive manufacturing (AM) can fabricate cellular and porous solids with exceptional strength-to-weight ratios or other performance properties. However, current design tools either (i) respect the outer surface but ignore mechanical loads or (ii) optimize lattice density but distort the shape. Traditional computational design methods also require tedious manual editing and long processing times.

In this project, we address these challenges by exploring modern generative diffusion models. Our approach conditions 3D diffusion models on both the shape’s surface and associated vector or scalar performance fields, enabling automatic generation of manufacturable, performance-aware microstructures. Bridging this gap promises lighter, stronger components for aerospace, biomedical scaffolds, and energy absorbers.

microRNAs (miRNAs) are short non-coding RNAs that regulate gene expression and play essential roles in development and disease. Their function depends on tightly regulated biogenesis, involving the processing of structured RNA precursors by enzymatic machinery. While sequence and secondary structure (2D) features have been studied extensively, they do not fully explain how cells distinguish miRNAs from other structured RNAs or determine processing efficiency. We hypothesize that 3D structural features more accurately reflect the physical and biochemical cues governing miRNA maturation.

This project aims to develop explainable graph neural network (GNN) models to learn the 3D structural determinants of miRNA processing. RNA structures will be represented as spatial graphs enriched with geometric and biochemical features. GNNs will be trained to classify precursors based on processing outcomes, and explainability techniques will be employed to identify key structural motifs driving model decisions. The goal is to mechanistically understand what makes an RNA molecule an optimal substrate for miRNA biogenesis. This work bridges computational modeling, machine learning, and RNA biology, contributing to the broader effort of applying explainable AI to molecular structure-function prediction.

Our project aims to develop practical and usable defenses against side-channel-based web privacy attacks. This class of attack can deanonymize web users and leak their private data by taking advantage of information revealed by web browsers and hardware-level system properties. Relatively few mitigations against these attacks have been adopted, due to concerns about degrading the user experience of browsers. However, this has not been supported with data about real users’ preferences and expectations. Our project intends to address this gap by studying side channel defenses from a human-centered point of view and developing novel side channel defenses based on these observations.

The problem of scheduling surgeries is challenging especially when simultaneously considering both emergency and elective cases and when surgery durations are uncertain. In this research we will model this problem as robust extensible bin packing with uncertain item size and bin sizes. Uncertain bin sizes reflect the residual work shift hours after performing emergency surgeries whose arrivals as well as durations are uncertain. Key aspects of the proposed project include: - Solving larger scale packing and scheduling problems to accommodate 2-12 week planning horizons of major hospitals. - Introducing model extensions and computational methods for addressing elective surgery scheduling in conjunction with emergency surgeries and other interruptions to routine hospital operations. - Extending and applying recent advances in bin packing research.

This project aims to enhance the scalability and performance of graph algorithms and Graph Neural Networks (GNNs) through innovations in computational infrastructure. Graphs are central to many scientific and technological domains, including AI, computational physics, and data science. However, as datasets grow, existing algorithms and hardware struggle to keep pace. Building on a previous NJIT-BGU collaboration, the proposed work targets two key challenges. First, it seeks to develop efficient, GPU-optimized graph coarsening techniques critical for multiscale methods in physics simulations and GNNs. By leveraging insights from algebraic multigrid (AMG) methods, the team aims to improve pooling operations in GNNs, enabling deeper and more scalable models. Second, the project introduces quaternion-based GNNs, which promise improved memory access patterns and performance through 4D data representations. This will involve both theoretical advancement and hardware-aware implementation. The research will span CPU, GPU, and novel Maverick chip architectures. The collaboration includes academic experts from NJIT and BGU and industry support from NextSilicon, offering a well-rounded approach to addressing the computational bottlenecks in large-scale graph processing.

The overarching goal of this project is to design robots that support humans in their daily lives in tasks such as in-home rehabilitation (e.g., post-surgery muscle training) and manipulation (e.g., assembling furniture). Thus, key technical capabilities such robots should posses include (i) understanding the current task context as well as human contexts (e.g., pain and comfort levels), and (ii) adapting their behavior between observation and assistance states so as to balance between their usefulness in these contexts and human's perception of their own autonomy. However, state-of-the-art frameworks for robot adaptation such as Reinforcement Learning pose several challenges (e.g., dealing with sparse rewards) and are relatively under-studied when humans are in the loop. Through this work, we identify key challenges in, and investigate solutions for using wearable data-driven implicit feedback in RL settings for achieving robot adaptation during human-robot collaboration (HRC).