Export
Nidhi Vakil
· 1 previous affiliation
- University of Massachusetts LowellSince 2020
- PhD student · Since 2020
Research Overview
Nidhi Vakil focuses on curriculum learning strategies for graph neural networks and text graphs, exploring how data scheduling and sample difficulty impact model training. Her work integrates textual semantics with graph topology to improve relation extraction and natural language processing tasks. Recently, she has investigated complexity-guided training paradigms that leverage model competence and spaced repetition to enhance generalization.
Research Trajectory
Nidhi Vakil’s research trajectory is defined by a consistent focus on optimizing the training dynamics of neural networks operating on structured data. As a lead author across her published work, she has developed sophisticated data scheduling techniques that move beyond standard monotonic training progressions, focusing instead on the nuanced relationship between data complexity and model learning.
In 2022, Vakil began publishing on the intersection of curriculum learning and relation extraction, contributing to the research area of Integrating Textual Semantics and Graph Topology. In her first-author paper Generic and Trend-aware Curriculum Learning for Relation Extraction, she proposed a model that treats entities as node pairs within text graphs. Rather than relying on static heuristics for difficulty, this approach incorporated sample-level loss trends to dynamically discriminate between easier and harder samples. This work aligns closely with the broader theme of The Linear Memory Bottleneck of Per-Sample State Tracking, demonstrating that tracking how a model interacts with individual data points over time yields a more robust estimation of sample difficulty and improves overall performance.
In 2023, Vakil expanded her focus to more comprehensive formalisms of graph complexity. She led the development of Complexity-Guided Curriculum Learning for Text Graphs, which introduced a novel data scheduler utilizing "spaced repetition." By building on established text and graph complexity metrics, this research demonstrated how tailored exposure to diverse examples can improve generalization while requiring less data. This approach directly engages with the research direction of The Intrinsic Difficulty of Data Samples, treating difficulty not merely as a function of the model's current state, but as an inherent property of the graph's structural and textual features.
Continuing this trajectory in 2023, Vakil addressed the limitations of using a single difficulty criterion in graph learning. In Curriculum Learning for Graph Neural Networks: A Multiview Competence-based Approach, she introduced a scheduling scheme that combines graph complexity formalisms with the model's real-time competence. By continuously evaluating both the intrinsic difficulty of the data and the evolving capability of the model, this multiview approach allows for highly adaptive training sequences. This methodology reflects advancements in the area of Non-Monotonic and Fluctuating Curriculum Learning, where training regimens are designed to fluctuate in difficulty to optimize the loss landscape for graph neural networks in language applications.
Current Directions
The most recent papers in the evidence indicate a sustained focus on advanced data scheduling and the evaluation of graph neural networks. Vakil’s current work investigates how multiple criteria of difficulty—specifically structural graph complexity combined with dynamic model competence—can be synthesized to create highly efficient training curricula.
This focus on the rigorous evaluation and structured processing of complex data naturally intersects with broader research themes visible in her recent trajectory, such as the area of Multimodal Graph Retrieval Augmented Generation. Furthermore, as her curriculum learning models require diverse and structurally complex data to accurately gauge model competence, her work touches upon the emerging research direction of Controllable Synthetic Generation for Graph Model Benchmarking. Ultimately, Vakil's recent directions point toward increasingly adaptive, data-efficient training paradigms that bridge the gap between linguistic semantics and graph-based structural reasoning.
Profile updated
Loading new papers…
Other researchers whose papers cluster into the same research directions — a proxy for overlapping technical interests and shared intellectual neighborhoods.
See something wrong? Report a mistake