2021 Research Prize Winners

1St. Prize: Simulating and Learning the Physical systems: Clebsch Gauge Fluids

Shuqi Yang (Computer Science), Advisor: Bo Zhu

Capturing and evolving the visually appealing vorticities is challenging for computer graphics and computational physics. Our research shows that such vivid and intricate vorticity evolution can be generated and preserved robustly in a numerical setting by solving a gauge transformation of Navier-Stokes equations. We devise our computational tool based on Clebsch wave functions, a critical mathematical expression widely used in quantum mechanics. Our method combines the expressive power of Clebsch wave functions to represent coherent vortical structures and the generality of gauge methods to accommodate different types of fluid. We showcased the efficacy of our computational approach by simulating a broad array of fluid phenomena, including leapfrogging, turbulent smoke, and surface tension flow.

Project page: https://y-sq.github.io/proj/clebsch_gauge_fluid/

Shuqi Yang, Shiying Xiong, Yaorui Zhang, Fan Feng, Jinyuan Liu, and Bo Zhu. Clebsch Gauge Fluid. ACM Transactions on Graphics (SIGGRAPH 2021), 40(4):99.

2nd. Prize: Between- subject Prediction Reveals a Shared Representational Geometry in the Rodent Hippocampus

Hung-tu Chen (Psychological & Brain Sciences), Advisor: Matthijs van der Meer

The hippocampus is a brain structure known for storing a map of episodic nature for each environment. In particular, neurons in  the hippocampus selectively fire in their preferred locations (place fields) when an animal runs through the environment. However, where hippocampal place cells have their fields is famously hard to predict: if you know how a given subject encodes location of environment A, that doesn't tell you much about how it encodes B.

We adapted a technique from human neuroimaging work (hyperalignment, inspired by Haxby Lab) enabling us to use how subject 1 encodes A and B (e.g. left and right arms of a maze), and how subject 2 encodes A, to predict how subject 2 encodes B.

Surprisingly, this cross prediction worked better than the within-subject controls we tried, and simulations suggest simple explanations such as correlated firing rates between A and B can be ruled out. Thus, we think between-subject prediction is a novel analysis approach that suggests an underlying regularity in how different places are mapped in the rodent hippocampus.

Chen, H. T., Manning, J. R., & van der Meer, M. A. (2020). Between-subject prediction reveals a shared representational geometry in the rodent hippocampus. bioRxiv

2nd. Prize: Mitigating Political Bias in Language Models through Reinforced Calibration

Rubio Liu (Computer Science), Advisor: Soroush Vosoughi

As more potentially-biased language models are adopted in AI applications, it is a growing concern that the political bias will be amplified if fairness is not taken into considering. In this work, we investigate the popular language model, GPT-2, to demonstrate what political bias is and how to mitigate such bias in generative language models. We describe metrics for measuring the bias and propose a reinforcement learning (RL) framework for mitigating political bias in the generated text. Requiring neither collecting extra data nor retraining the model from scratch, our method is especially meaningful in realistic settings. In empirical experiments on three attributes sensitive to political bias (gender, location, and topic), our methods reduced bias according to both our metrics and human evaluation, while maintaining readability and semantic coherence.

3rd Prize: Two-Step Training Deep Learning Framework for Computational Imaging without Physics Prior

Rubio Shang (Thayer School of Engineering), Advisor: Geoffrey P. Luke

Computational imaging is an emerging field where the hardware requirements of an imaging system are relaxed by incorporating algorithmic reconstruction of the image. A critical challenge in computational imaging is that imperfections in the idealized mathematical model of the imaging system degrade reconstruction performance. Furthermore, many imaging inverse problems are ill-posed, which leads to suboptimal image reconstruction. We developed a two-step training deep learning framework to solve these challenges. It directly learns the inverse model from the data in its first step of training and applies regularization constraints in its second. Overall, it is a flexible framework for diverse imaging systems.

Ruibo Shang, Kevin Hoffer-Hawlik, Fei Wang, Guohai Situ, and Geoffrey P. Luke, "Two-step training deep learning framework for computational imaging without physics priors," Opt. Express 29, 15239-15254 (2021)

https://www.osapublishing.org/oe/fulltext.cfm?uri=oe-29-10-15239&id=450666

1St. Prize: Sparse Symplectically Integrated Neural Networks

Daniel DiPietro (Computer Science), Advisor: Bo Zhu

Our work seeks to answer the following question: given the historical data of some physical dynamical system, how can we not only predict its future states, but also discern its underlying governing equations? To do this, we introduce Sparse Symplectically Integrated Neural Networks (SSINNs), a novel machine learning architecture that combines fourth-order symplectic integration with a learned parameterization of the Hamiltonian obtained using sparse regression through a mathematically elegant function space. This allows for interpretable models that incorporate symplectic inductive biases and have low memory requirements. SSINNs succeed in learning governing equations from as few as 200 noisy data points and often outperform current state-of-the-art methods by an order of magnitude.

DiPietro, D. M., Xiong, S., & Zhu, B. (2020). Sparse Symplectically Integrated Neural Networks. Advances in Neural Information Processing Systems 33, 6074-6085.

2nd. Prize: Fine-Grained Detections of Hate Speech Using BERToxic

Yakoob Khan (Computer Science), Advisor: Soroush Vosoughi

The rise of hate speech is a challenging problem for online platforms. To address this issue, we propose BERToxic, a system that fine-tunes a pre-trained BERT model to locate toxic text spans in a given text and utilizes additional post-processing steps to refine the prediction boundaries. Through experiments, we show that the post-processing steps improve the performance of our model by 4.16% on the test set. We also studied the effects of data augmentation and ensemble modeling strategies on our system. Our model significantly outperformed strong baselines and achieved an F1-score of 0.683, placing our system competitively in a hate speech detection competition.

Yakoob Khan, Weicheng Ma, and Soroush Vosoughi. (2021). Lone Pine at SemEval-2021 Task 5: Fine-Grained Detection of Hate Speech Using BERToxic. In Proceedings of the 15th International Workshop on Semantic Evaluation (SemEval-2021). To appear.

2nd. Prize: Impacts of Mining & Dam-Building on the Sediment Flux of the Maderia River in the Amazon River Basin: A Big Data Approach

Shannon Sartain (Earth Science), Advisor: Carl Renshaw

The Amazon River Basin is increasingly subject to anthropogenic impacts, like mining and dam-building, which directly alter natural sediment cycles. To assess the magnitude and spatial extent of these impacts on sediment flux, we construct 3000-kilometer sediment flux profiles over 35 years of the Madeira River, the largest tributary to the Amazon with respect to sediment delivery. We use two models based on long-term, remotely collected data: one to estimate daily discharge from satellite-generated precipitation and one to estimate suspended sediment concentration from satellite imagery. We find that the magnitude of impact of mining relative to pre-disturbance sediment cycles is minimal compared to that of dams. However, incoming tributaries, which replenish water and sediment, ameliorate the impacts of both downstream. Our results illuminate the potential for impacts to persist in future years and further downstream if key replenishing tributaries are dammed or otherwise altered by humans.