A5102795202- Advanced Optical Imaging Technologies
- Advanced optical system design
- Semiconductor Lasers and Optical Devices
- Luminescence Properties of Advanced Materials
- Color Science and Applications
- Impact of Light on Environment and Health
- Solid State Laser Technologies
- Retinal Diseases and Treatments
- Surface Roughness and Optical Measurements
- Optical measurement and interference techniques
- Topic Modeling
- Retinal Imaging and Analysis
- Photonic and Optical Devices
- Infrared Target Detection Methodologies
- Ocular and Laser Science Research
- Spacecraft Design and Technology
- Optical Systems and Laser Technology
- Glaucoma and retinal disorders
- Advanced Semiconductor Detectors and Materials
- Natural Language Processing Techniques
- Cancer Diagnosis and Treatment
- Diet, Metabolism, and Disease
- Liver physiology and pathology
- Speech and dialogue systems
- Advanced Optical Sensing Technologies
Harvard University
2024
Icahn School of Medicine at Mount Sinai
2022-2023
Washington University in St. Louis
2022
United Christian Hospital
2020-2021
Vanderbilt University
2021
InVivo Analytics (United States)
2018
Queen Elizabeth Hospital
2012-2018
California State University Los Angeles
2012
TRW Automotive (United States)
1983
Language models show a surprising range of capabilities, but the source their apparent competence is unclear. Do these networks just memorize collection surface statistics, or do they rely on internal representations process that generates sequences see? We investigate this question by applying variant GPT model to task predicting legal moves in simple board game, Othello. Although network has no priori knowledge game its rules, we uncover evidence an emergent nonlinear representation state....
We introduce Inference-Time Intervention (ITI), a technique designed to enhance the "truthfulness" of large language models (LLMs). ITI operates by shifting model activations during inference, following set directions across limited number attention heads. This intervention significantly improves performance LLaMA on TruthfulQA benchmark. On an instruction-finetuned called Alpaca, its truthfulness from 32.5% 65.1%. identify tradeoff between and helpfulness demonstrate how balance it tuning...
Diabetic macular edema (DME) is the primary cause of vision loss among individuals with diabetes mellitus (DM). We developed, validated, and tested a deep learning (DL) system for classifying DME using images from three common commercially available optical coherence tomography (OCT) devices.We trained validated two versions multitask convolution neural network (CNN) to classify (center-involved [CI-DME], non-CI-DME, or absence DME) three-dimensional (3D) volume scans 2D B-scans,...
Readers find text difficult to consume for many reasons. Summarization can address some of these difficulties, but introduce others, such as omitting, misrepresenting, or hallucinating information, which be hard a reader notice. One approach addressing this problem is instead modify how the original rendered make important information more salient. We Grammar-Preserving Text Saliency Modulation (GP-TSM), rendering method with novel means identifying what de-emphasize. Specifically, GP-TSM...
Prompting is a standard tool for customizing language-model chatbots, enabling them to take on specific "persona". An implicit assumption in the use of prompts that they will be stable, so chatbot continue generate text according stipulated persona duration conversation. We propose quantitative benchmark test this assumption, evaluating stability via self-chats between two personalized chatbots. Testing popular models like LLaMA2-chat-70B, we reveal significant drift within eight rounds...
<a><b>Objective:</b></a> Diabetic macular edema (DME) is the primary cause of vision loss among individuals with diabetes mellitus (DM). We developed, validated, and tested a deep-learning (DL) system for classifying DME using images from three common commercially available optical coherence tomography (OCT) devices. <p><b>Research Design Methods:</b> trained validated two versions multi-task convolution neural network (CNN) to classify...
We demonstrate the generation of a laser hotspot by laser-excited single crystal phosphor (SCP) converter layer in white LED for producing wide field view (FOV) and high enhanced central brightness (ECB) an adaptive-driving-beam (ADB) headlight. The SCP shows better thermal quench behavior than other layers because higher melting temperature 1940 °C larger conductivity 11 W/m°C. advantage this study is optical design ADB headlight, which generation, mechanism, characterization are...
A novel laser-assisted LED for adaptive-driving-beam (ADB) headlights employing an ultra-reliable Ce 3+ : YAG-based single crystal phosphor (SCP)-converter layer use in autonomous vehicles is demonstrated. The SCP fabricated at a high-temperature of 1,940°C exhibited better thermal stability than other phosphor-converter materials, evidenced by aging test. high-beam pattern the ADB measured luminous intensity 88,436 cd 0°, 69,393 ± 2.5°, and 42,942 5°, which well satisfies ECE R112 class B...
Laser excited phosphor has been used to excite material, producing high intensity light output with smaller etendue than that of LEDs the same long lifetime. But due laser light, organic binder burns at low power, which requires be deposited on a rotating wheel in practical applications. Phosphor inorganic binders, commonly known as ceramic phosphor, other hand, does not burn, but efficiency goes down temperature up under power excitation. This paper describes cooling schemes sealed chambers...
A general theory on light-pipe/tunnel-based recycling schemes, along with some experimental data obtained from LED illuminators, is presented in this paper. Light classified as angular, spatial, or mixed based the angular filtering characteristics of process. The principles behind three schemes are analyzed and results at end.
Abstract LEDs are still not bright enough, compared to arc lamps, for high power projector applications. For applications, multiple required increase the total light output with same amount of in etendue. Even use larger imaging panels etendue values, LED is sufficient cinema projection With recycling technology, can be reduced increased brightness. This paper describes two systems an array 7 recycling, such that combined matches standard imagers The first system uses a single white up 9,500...
Readers find text difficult to consume for many reasons. Summarization can address some of these difficulties, but introduce others, such as omitting, misrepresenting, or hallucinating information, which be hard a reader notice. One approach addressing this problem is instead modify how the original rendered make important information more salient. We Grammar-Preserving Text Saliency Modulation (GP-TSM), rendering method with novel means identifying what de-emphasize. Specifically, GP-TSM...
We present an approach called Q-probing to adapt a pre-trained language model maximize task-specific reward function. At high level, sits between heavier approaches such as finetuning and lighter few shot prompting, but can also be combined with either. The idea is learn simple linear function on model's embedding space that used reweight candidate completions. theoretically show this sampling procedure equivalent KL-constrained maximization of the Q-probe number samples increases. To train...
Conversational LLMs function as black box systems, leaving users guessing about why they see the output do. This lack of transparency is potentially problematic, especially given concerns around bias and truthfulness. To address this issue, we present an end-to-end prototype-connecting interpretability techniques with user experience design-that seeks to make chatbots more transparent. We begin by showing evidence that a prominent open-source LLM has "user model": examining internal state...
We present an approach called Dialogue Action Tokens (DAT) that adapts language model agents to plan goal-directed dialogues. The core idea is treat each utterance as action, thereby converting dialogues into games where existing approaches such reinforcement learning can be applied. Specifically, we freeze a pretrained and train small planner predicts continuous action vector, used for controlled generation in round. This design avoids the problem of degradation under reward optimization....
The Wavien dual paraboloid reflector system is an imaging having a 1:1 magnification characteristic that allows focusing the light from arc into target pipe with practically no change in etendue. A retro-reflector used to increase brightness by reflecting part of back through itself. Due ray distortion created envelop burner, especially for UHP lamps where they are very thick and uneven, reflected images aberrated reducing coupling efficiency. In this paper, non-spherical correct image...
Abstract A general theory on light‐pipe/tunnel‐based recycling schemes, along with some experimental data obtained from LED illuminators, is presented in this paper. Light classified as angular, spatial, or mixed based the angular filtering characteristics of process. The principles behind three schemes are analyzed and results at end.
Abstract With the advancements in LED development, it is now possible to have projectors small enough be integrated into cell phones. Similar phone cameras, order a high adoption rate, cost low, which requires low component count and simple construction. To achieve cost, most important factor use imager chip, increase illumination requirements. Recycling needed match etendue that of chip. This paper presents an approach with recycling projection architecture using waveguides such whole...
The rapid development of high power visible LEDs has prompted an increase in interest the projection display industry. Although level not achieved sufficient for large displays, it is smaller displays. It also expected that near future, there will be significant improvements LED output and market may open up. In this paper, a based illumination engine presented which suitable using array whose coupled into single light pipe with minimum loss brightness. basic building block consists chip...
The LiDAR-embedded smart laser headlight module employing a single digital micromirror device is demonstrated. This novel scheme enables reduction in optical components and space requirements for use the next-generation high-performance autonomous vehicles.