skip to main content

Attention:

The NSF Public Access Repository (NSF-PAR) system and access will be unavailable from 11:00 PM ET on Thursday, May 23 until 2:00 AM ET on Friday, May 24 due to maintenance. We apologize for the inconvenience.


Search for: All records

Creators/Authors contains: "Li, Y"

Note: When clicking on a Digital Object Identifier (DOI) number, you will be taken to an external site maintained by the publisher. Some full text articles may not yet be available without a charge during the embargo (administrative interval).
What is a DOI Number?

Some links on this page may take you to non-federal websites. Their policies may differ from this site.

  1. Transformer interpretability aims to understand the algorithm implemented by a learned Transformer by examining various aspects of the model, such as the weight matrices or the attention patterns. In this work, through a combination of theoretical results and carefully controlled experiments on synthetic data, we take a critical view of methods that exclusively focus on individual parts of the model, rather than consider the network as a whole. We consider a simple synthetic setup of learning a (bounded) Dyck language. Theoretically, we show that the set of models that (exactly or approximately) solve this task satisfy a structural characterization derived from ideas in formal languages (the pumping lemma). We use this characterization to show that the set of optima is qualitatively rich; in particular, the attention pattern of a single layer can be “nearly randomized”, while preserving the functionality of the network. We also show via extensive experiments that these constructions are not merely a theoretical artifact: even with severe constraints to the architecture of the model, vastly different solutions can be reached via standard training. Thus, interpretability claims based on inspecting individual heads or weight matrices in the Transformer can be misleading. 
    more » « less
    Free, publicly-accessible full text available December 10, 2024
  2. Free, publicly-accessible full text available March 1, 2025
  3. The estimation of malaria parasite migration can play a vital role in informing elimination strategies by pinpointing regions with higher parasite migration that act as transmission sources, and that could be the focus of elimination interventions. Gene flow simulation methods such as Estimated Effective Migration Surfaces (EEMS) and Migration and Population-Size Surfaces (MAPS) use a Markov Chain Monte Carlo simulation-based approach to visualize a species' migration and diversity. These methods utilize georeferenced genomic data and present output in the form of migration contour maps. Despite their potential, there is uncertainty in EEMS and MAPS outputs when sampling locations are sparse - an aspect that remains under-explored in current research. We present a framework designed to systematically assess the impact of sample locations and sample size on migration contours in gene flow simulations that goes beyond the posterior probability map available in EEMS. We test our framework using publicly available genomic data collected from Cambodia and border regions of Thailand, Vietnam, and Laos during 2008-2013. The methodology leverages kernel density estimation and topological skeletons in conjunction with other spatial analysis methods to quantify the impact of sparse sample locations on gene flow simulations. Multiple sample resolutions were tested against a baseline resolution, and the findings highlight how migration contours vary with sampling resolution and how our approach can be applied to guide the production and mapping of reliable migration contours. Our research provides valuable insights about both the reliability and precision of model outputs when employing gene flow simulation techniques e.g., EEMS and MAPS, to estimate malaria parasite migration. The findings revealed that by employing our approach, we were able to maintain approximately 67% consistency between the contours and the reference dataset, even when utilizing only half of the sample locations. This knowledge will improve both the reliability and precision of these model outputs in future studies. 
    more » « less
  4. Free, publicly-accessible full text available December 1, 2024
  5. Abstract

    Across varied environments, meandering channels evolve through a common morphodynamic feedback: the sinuous channel shape causes spatial variations in boundary shear stress, which cause lateral migration rates to vary along a meander bend and change the shape of the channel. This feedback is embedded in all conceptual models of meandering channel migration, and in numerical models, it occurs over an explicit timescale (i.e., the model time step). However, the sensitivity of modeled channel trajectory to the time step is unknown. In numerical experiments using a curvature‐driven model of channel migration, we find that channel trajectories are consistent over time if the channel migrates ≤10% of the channel width over the feedback timescale. In contrast, channel trajectories diverge if the time step causes migration to exceed this threshold, due to the instability in the co‐evolution of channel curvature and migration rate. The divergence of channel trajectories accumulates with the total run time. Application to hindcasting of channel migration for 10 natural rivers from the continental US and the Amazon River basin shows that the sensitivity of modeled channel trajectories to the time step is greatest at low (near‐unity) channel sinuosity. A time step exceeding the criterion causes over‐prediction of the width of the channel belt developed over millennial timescales. These findings establish a geometric constraint for predicting channel migration in landscape evolution models for lowland alluvial rivers, upland channels coupled to hillslopes and submarine channels shaped by turbidity currents, over timescales from years to millennia.

     
    more » « less
  6. Free, publicly-accessible full text available August 28, 2024
  7. Free, publicly-accessible full text available August 21, 2024