arXiv subjects

Alan Lee

Publications and source records attributed to Alan Lee.

10 recordsLinked to original sources

arXiv2026

Automated Extraction of Unstructured Post-SBRT Toxicity Data from Radiology Reports Using Large Language Models

We evaluated the viability of using a Large Language Model (LLM) to extract patient-specific specific toxicity and progression outcomes from unstructured radiology reports. We retrospectively extracted 160 follow-up CT and PET/CT electronic medical record notes for patients treated with lung stereotactic body radiotherapy (SBRT) at our institution from January 2017 through December 2023. Using the Llama 3.3-70-B-Instruct LLM, we engineered prompts to extract four clinical endpoints from each radiology report: locoregional progression, distant progression, radiation-induced fibrosis, and radiation-induced rib fractures. Progression endpoints were classified as yes, no, or maybe, while fibrosis and rib fractures were binary (yes or no). Ground truth labels were defined using two-grader consensus for the 60-note training set, used for prompt development, and a three-grader majority vote for the 100-note test set. LLM performance was evaluated using sensitivity, specificity, and accuracy. As detailed by our evaluation metrics, the strong performance of our methods demonstrates the viability of using prompt-engineered LLMs to extract radiation-toxicities and progression classification from radiology reports.

physics.med-ph↗

arXiv2025

Flexible-AR display for near-eye operations

We propose a new technique to fabricate flexible-near-field Argument-Reality (AR) display using modular-molds. A near-eye flexible-AR-display is fabricated based on parameters extracted from simulations. Our AR-display successfully reconstructed images and videos from a light-engine. It opens a new approach to fabricate flexible-near-field AR display with good physical stress and collision-resilience.

physics.optics↗

arXiv2025

Token-Efficient RL for LLM Reasoning

We propose reinforcement learning (RL) strategies tailored for reasoning in large language models (LLMs) under strict memory and compute limits, with a particular focus on compatibility with LoRA fine-tuning. Building on early policy gradient methods with baseline subtraction, we design critic-free methods that operate on a small, informative subset of output tokens to reduce memory usage and stabilize training. We introduce S-GRPO, a stochastic variant of Group Relative Policy Optimization, and T-SPMO, a token-level prefix matching approach for fine-grained credit assignment. Applied to Qwen2-1.5B, our methods raise accuracy on the SVAMP benchmark from 46% to over 70% and show strong performance on multi-digit multiplication. Surprisingly, full-token GRPO under LoRA fails to improve over the base model, suggesting that selective token-level optimization may act as an implicit regularizer in low-parameter training regimes.

cs.LG↗

arXiv2024

Design and Fabrication of a Low-cost Liquid Optical Waveguide for Augmented Reality

The complexities of fabrication techniques and the demand for high precision have posed significant challenges in the mass production of augmented reality (AR) waveguide combiners. Leveraging the capabilities of Polyjet 3D printing techniques, we have developed a cost-effective method for fabricating liquid geometric waveguide combiners for AR applications, using silicone oil as the medium. During the design phase, we optimized the structure of the waveguide combiner to facilitate easier fabrication. Our proposed method simplifies the production process by removing the need for complicated steps like dicing, layer bonding, and polishing, which are usually involved in traditional manufacturing techniques. We conducted optical simulations and developed a prototype using our patented fabrication method, which successfully demonstrated the integration of virtual images with the real-world environment, thereby confirming its feasibility and potential for cost-effective mass production.

physics.optics↗

arXiv2024

3D Printed Waveguide for Augmented Reality

Mass production of augmented reality (AR) waveguides has been challenging due to the intricate nature of the fabrication technique and the high precision required for its optical characteristics. In this paper, we have presented a novel and low-cost approach for fabricating geometric optical waveguides designed for AR applications utilizing 3D printing techniques. To strike a balance between optical performance and fabrication feasibility, we have optimized the conventional geometric waveguide design to facilitate easier fabrication. It is worth noting that our proposed method does not require molding, dicing, and post-surface polishing after printing. A prototype based on this method has been successfully fabricated, showing the immersion between the virtual image and the real-world scene. The proposed method has great potential for adaptation to mass production in various AR applications.

physics.optics↗

arXiv2023

Compact Metasurface Terahertz Spectrometer

The electromagnetic spectrum in the terahertz frequency region is of significant importance for understanding the formation and evolution of galaxies and stars throughout the history of the universe and the process of planet formation. Within the star forming clouds the constituent atoms and molecules are excited to produce characteristic emission and absorption lines, many of which happen at the terahertz frequencies. Thus, detecting the spectral signatures as unique fingerprints of molecules and atoms require terahertz spectrometers, which need to be operated in a space observatory because of the water vapor absorption in the earth atmosphere. However, current terahertz spectrometers face several challenges that limit their performances and applications, including a low resolution, limited bandwidth, large volume, and complexity. In this paper, we address the last two issues by demonstrating a concept of a compact terahertz spectrometer using metasurface. We start by modelling, designing, and fabricating a metasurface, aiming to optimize its performance within a band from 1.7 to 2.5 THz. Next, we make use of an array of quantum cascade lasers that operate at slightly different frequencies around 2.1 THz to validate the performance of the spectrometer. Finally, we apply the spectrum inversion method to analyse the measured data to confirm a resolution R of at least 273. Our results demonstrated a miniaturized terahertz spectrometer concept successfully.

physics.optics↗

arXiv2022

Connectedness in Friends-and-Strangers Graphs of Spiders and Complements

Let $X$ and $Y$ be two graphs with vertex set $[n]$. Their friends-and-strangers graph $\mathsf{FS}(X,Y)$ is a graph with vertices corresponding to elements of the group $S_n$, and two permutations $\sigma$ and $\sigma'$ are adjacent if they are separated by a transposition $\{a,b\}$ such that $a$ and $b$ are adjacent in $X$ and $\sigma(a)$ and $\sigma(b)$ are adjacent in $Y$. Specific friends-and-strangers graphs such as $\mathsf{FS}(\mathsf{Path}_n,Y)$ and $\mathsf{FS}(\mathsf{Cycle}_n,Y)$ have been researched, and their connected components have been enumerated using various equivalence relations such as double-flip equivalence. A spider graph is a collection of path graphs that are all connected to a single center point. In this paper, we delve deeper into the question of when $\mathsf{FS}(X,Y)$ is connected when $X$ is a spider and $Y$ is the complement of a spider or a tadpole.

math.CO↗

arXiv2022

Connectedness and Cycle Spaces of Friends-and-Strangers Graphs

If $X=(V(X),E(X))$ and $Y=(V(Y),E(Y))$ are $n$-vertex graphs, then their friends-and-strangers graph $\mathsf{FS}(X,Y)$ is the graph whose vertices are the bijections from $V(X)$ to $V(Y)$ in which two bijections $\sigma$ and $\sigma'$ are adjacent if and only if there is an edge $\{a,b\}\in E(X)$ such that $\{\sigma(a),\sigma(b)\}\in E(Y)$ and $\sigma'=\sigma\circ (a\,\,b)$, where $(a\,\,b)$ is the permutation of $V(X)$ that swaps $a$ and $b$. We prove general theorems that provide necessary and/or sufficient conditions for $\mathsf{FS}(X,Y)$ to be connected. As a corollary, we obtain a complete characterization of the graphs $Y$ such that $\mathsf{FS}(\mathsf{Dand}_{k,n},Y)$ is connected, where $\mathsf{Dand}_{k,n}$ is a dandelion graph; this substantially generalizes a theorem of the first author and Kravitz in the case $k=3$. For specific choices of $Y$, we characterize the spider graphs $X$ such that $\mathsf{FS}(X,Y)$ is connected. In a different vein, we study the cycle spaces of friends-and-strangers graphs. Naatz proved that if $X$ is a path graph, then the cycle space of $\mathsf{FS}(X,Y)$ is spanned by $4$-cycles and $6$-cycles; we show that the same statement holds when $X$ is a cycle and $Y$ has domination number at least $3$. When $X$ is a cycle and $Y$ has domination number at least $2$, our proof sheds light on how walks in $\mathsf{FS}(X,Y)$ behave under certain Coxeter moves.

math.CO↗

arXiv2021

Moving outside the lab: The viability of conducting sensorimotor learning studies online

Collecting data online via crowdsourcing platforms has proven to be a very efficient way to recruit a large and diverse sample. Studies of motor learning, however, have been largely confined to the lab due to the need for special equipment to record movement kinematics and, as such, are typically only accessible to specific participants (e.g., college students). As a first foray to make motor learning studies accessible to a larger and more diverse audience, we developed an online, web-based platform (OnPoint) to collect kinematic data, serving as a template for researchers to create their own online sensorimotor control and learning experiments. As a proof-of-concept, we asked if fundamental motor learning phenomena discovered in the lab could be replicated online. In a series of three experiments, we observed a close correspondence between the results obtained online with those previously reported from research conducted in the laboratory. This web-based platform paired with online crowdsourcing can serve as a powerful new method for the study of motor control and learning.

q-bio.NC↗

arXiv2012

Reparametrization of the least favorable submodel in semi-parametric multisample models

The method of estimation in Scott and Wild (Biometrika 84 (1997) 57--71 and J. Statist. Plann. Inference 96 (2001) 3--27) uses a reparametrization of the profile likelihood that often reduces the computation times dramatically. Showing the efficiency of estimators for this method has been a challenging problem. In this paper, we try to solve the problem by investigating conditions under which the efficient score function and the efficient information matrix can be expressed in terms of the parameters in the reparametrized model.

math.ST↗