23 research outputs found
Bei ji qian jin yao fang: 30 juan. v.1
No full text孫思邈撰 ; 林億校.框21.9x15.1公分, 13行23字, 白口, 左右雙邊, 雙黑魚尾, 版心中鐫書名, 卷次.綫裝, 2函.江戶醫學北宋栞本景摹開雕.Sun Simiao zhuan ; Lin Yi jiao.Kuang 21.9 x 15.1 gong fen, 13 hang 23 zi, bai kou, zuo you shuang bian, shuang hei yu wei, ban xin zhong juan shu ming, juan ci.Xian zhuang, 2 han.Jianghu yi xue BeiSong kan ben ying mo kai diao
Bei ji qian jin yao fang: san shi juan. v.1
No full text[孫思邈撰] ; 林億...[et al.]校正 .綫裝.框20.9x14.7公分, 13行23字. 白口, 間中有黑口, 左右雙邊, 順黑魚尾. 版心中鐫"千金要方"及卷次, 下鐫葉次及刻工.書名背頁刻"江戶醫學影北宋本, 光緖戊寅[1878]夏五購自東瀛, 印於上海, 長洲黃學熙記".目錄卷端下刻"金澤文庫" ; 書末有牌記刻"嘉永紀元江戶醫學北宋槧本景摹開雕"此書為日本嘉永2年[1849]江戶醫學據北宋本影刻, 後黃學熙得到書版, 重新刷印.《中國中醫古籍總目》(03309)著錄.鈐"莊兆祥印", "莊兆祥".Xian zhuang.Kuang 20.9 x 14.7 gong fen, 13 hang 23 zi. Bai kou, jian zhong you hei kou, zuo you shuang bian, shun hei yu wei. Ban xin zhong juan "Qian jin yao fang" ji juan ci, xia juan ye ci ji ke gong.Detailed notes in vernacular field only.Detailed notes in vernacular field only.Detailed notes in vernacular field only.Detailed notes in vernacular field only.[Sun Simiao zhuan] ; Lin Yi ...[et al.] jiao zheng .Qian "Zhuang Zhaoxiang yin", "Zhuang Zhaoxiang"
On Training, Inference, and Sample Efficiencies of Language Models
No full textLarge language models have demonstrated superior performance in various natural language processing tasks such as machine translation, natural language understanding, and natural language generation. However, despite the recent developments, language models still face critical challenges. In this thesis, we investigate efficient training and inference algorithms. We also investigate the sample efficiency of training language models.
In Chapter 2, we improve training efficiency of sparsely activated models by proposing a novel Mixture-of-Experts architecture. In Chapter 3, we propose state space augmented Transformer models, facilitating efficient modeling of long sequences. In Chapter 4, we target for inference efficiency of pre-trained language models. Specifically, we propose a knowledge distillation algorithm which adapts a pre-trained model into a Mixture-of-Experts model. In Chapter 5, we design a label efficient self-training algorithm. Specifically, we integrate differentiable teacher models into the conventional teacher-student self-training framework.Ph.D
Evoke: Evoking Critical Thinking Abilities in LLMs via Reviewer-Author Prompt Editing
No full textLarge language models (LLMs) have made impressive progress in natural
language processing. These models rely on proper human instructions (or
prompts) to generate suitable responses. However, the potential of LLMs are not
fully harnessed by commonly-used prompting methods: many human-in-the-loop
algorithms employ ad-hoc procedures for prompt selection; while auto prompt
generation approaches are essentially searching all possible prompts randomly
and inefficiently. We propose Evoke, an automatic prompt refinement framework.
In Evoke, there are two instances of a same LLM: one as a reviewer
(LLM-Reviewer), it scores the current prompt; the other as an author
(LLM-Author), it edits the prompt by considering the edit history and the
reviewer's feedback. Such an author-reviewer feedback loop ensures that the
prompt is refined in each iteration. We further aggregate a data selection
approach to Evoke, where only the hard samples are exposed to the LLM. The hard
samples are more important because the LLM can develop deeper understanding of
the tasks out of them, while the model may already know how to solve the easier
cases. Experimental results show that Evoke significantly outperforms existing
methods. For instance, in the challenging task of logical fallacy detection,
Evoke scores above 80, while all other baseline methods struggle to reach 20
Design and control of multifunctional sorting and training platform based on PLC control
No full textNon-Invasive sleep apnea detection using microphone technology
No full textDespite advancements in device and machine learning programs for diagnosing obstructive sleep apnea (OSA), widespread adoption is hindered by key implementation barriers, including the need for more accurate and non-invasive diagnostic tools. Various methods, such as ECG, EEG, SpO2, and respiratory signals, are used for detection, but limitations in accuracy persist. This thesis focuses on developing a machine learning algorithm using Subspace KNN classification of respiratory signals to detect different breathing patterns in individuals with and without OSA. The research also explores optimization strategies for hardware components, using Arduino and PCB boards, to enhance data collection over multiple nights. By exploring respiratory-based approaches for OSA detection, this thesis aims to provide a more targeted and specific method to detect apnea events, enhance accuracy and efficiency in detection, and contribute to the advancement of a more user-friendly and accessible solutionM.S.Includes bibliographical reference
Development of a method to measure skin tone using a hyperspectral camera
No full textCurrent methods for quantifying skin tone lack standardization and tend to be subjective, often relying on human judgement. Research studies and clinical practice would both benefit from more consistent and objective approaches. An object’s color arises from its reflectance spectrum, which can be measured at each pixel value in an image by hyperspectral imaging. A hyperspectral camera was used to assess color samples and skin tones, examining its accuracy and precision in color measurement. A setup with broadband white light illumination was assembled with a hyperspectral camera acquiring images at 300 wavelength bands between 400-1000 nm. The system was used to image standard targets with known colorimetric values and then used to evaluate skin in human subjects. This study also compared the human perception of skin color with the hyperspectral camera’s measurements. The hyperspectral camera demonstrates a mean difference between measured and ground truth L*a*b* colors over the 24 patches of a Calibrite ColorChecker target of ΔE = 4.99. The hyperspectral camera showed lower variation (standard deviation) in human skin tone classification than a group of five human observers. This thesis serves to highlight important concepts in hyperspectral imaging and demonstrates the potential for this technology for improving methods of skin color evaluation.M.S.Includes bibliographical reference
Development of an optical imaging platform for assessment of tooth color and whiteness
No full textThe two most common devices to measure tooth color in dental applications are shade guides (a visual comparison) and colorimeters (a quantitative comparison). Shade guides are subjective, difficult to convert into a useful digital file type, and metrics cannot be easily converted to a standard color scale to quantify whiteness. Colorimeters can provide quantitative color measurements, but usually only at a single location on a sample. Digital color cameras offer the potential to achieve quantification of sample color mapped across an entire field of view but typically provide only three overlapping red / green / blue spectral channels. This project aims to create an optical imaging system using a new polarization-sensitive color camera to perform color and whiteness index measurements using the CIELAB color space. Values obtained from the camera setup were compared to measurements obtained from a commercial colorimeter (Optishade). The system was calibrated and validated using standard test objects (ColorChecker CG1 and VITA teeth), then tested using six bovine teeth which underwent staining and whitening.
After the bovine teeth were stained by immersion in coffee, there was a mean decrease in whiteness of 3.80% and mean change in L*a*b* color of 3.16% (n = 6). After whitening with a 30% hydrogen peroxide solution, there was a mean increase in whiteness of 5.98% and mean change in L*a*b* color of 4.30%, relative to tooth color after the staining experiment. The polarization camera system gave a mean relative error of 4.91% in L* value, 30.95% in a* value, and 8.21% in b* value when comparing the staining experiment results to the Optishade values. The polarization camera system gave mean relative errors of 7.57% in L* value, 20.37% in a* value, and 6.67% in b* value when comparing the post-whitening color to the Optishade values.
In summary, the methods developed in this project provide a foundation for using a new polarization-sensitive color camera for quantitative spatial mapping of tooth color and whiteness. Future studies may focus on decreasing errors in the measured L*a*b* values and using the camera’s different polarization images, which can carry information about light returning from superficial and deep layers of the tooth.M.S.Includes bibliographical reference
Electrostatic phenomena associated with peeling tape
No full textTriboelectrification or contact-electrification is a widely documented process that describes the generation of electrical charge from contacting and separating surfaces. Though it is not well defined, and its mechanisms remain widely debated despite decades of speculation and study, there are many examples of this behavior. As early as Picard’s observations of flashes of light from mercury rolling from a barometer, we have been aware of intriguing interactions in rubbing surfaces. It is responsible for the shock experienced when removing a wool sweater and the fascinating volcanic lightning made possible through granular collisions in ash plumes. In this dissertation, we discuss three projects investigating phenomena associated with triboelectricity. The first chapter examines the relationship between stick-slip events and triboelectricity in separating surfaces. Using the simple act of peeling adhesive tape, we establish protocols for visualizing stick-slip and charge patterns using electrically charged toners. We identify reproducible trends in the patterns on the tape and the substrate from which it was peeled, and by varying the substrate materials, we show that peeling tape from different materials (e.g., polymethylmethacrylate and polytetrafluoroethylene) may pre-determine certain characteristics by generating specific electrostatic and stick-slip patterns. This method may be valuable for classifying surfaces in industry or laboratory experiments based on their stick-slip signatures. In the second chapter, we examine complex electrostatic patterns from mated surfaces, including the complementary but distinct differences between surfaces. We classify notable signatures achieved with this technique. We discuss a computational model in which particles interact with each other and a charged surface. We show that varying the particles’ electrical polarity will replicate the two general categories of patterns we see in experimental surface separation. We directly compare the results of the model with experimental findings. We contextualize our simulation and experimental results with questions that remain unanswered and with future steps. In the last chapter, we evaluate three-dimensional structures formed from electrostatic interactions in granular materials. In micro-gravity environments, where electrostatic interactions dominate over gravity, granular tendrils have been identified, similar to those found in highly charged systems in laboratory experiments. We compare debris attracted to equipment on Mars in 2022 to formations composed of toner to motivate this work, adding to it chains of granular materials formed on triboelectrically charged adhesive tape.Ph.D.Includes bibliographical reference
Self-Training with Differentiable Teacher
No full textSelf-training achieves enormous success in various semi-supervised and
weakly-supervised learning tasks. The method can be interpreted as a
teacher-student framework, where the teacher generates pseudo-labels, and the
student makes predictions. The two models are updated alternatingly. However,
such a straightforward alternating update rule leads to training instability.
This is because a small change in the teacher may result in a significant
change in the student. To address this issue, we propose DRIFT, short for
differentiable self-training, that treats teacher-student as a Stackelberg
game. In this game, a leader is always in a more advantageous position than a
follower. In self-training, the student contributes to the prediction
performance, and the teacher controls the training process by generating
pseudo-labels. Therefore, we treat the student as the leader and the teacher as
the follower. The leader procures its advantage by acknowledging the follower's
strategy, which involves differentiable pseudo-labels and differentiable sample
weights. Consequently, the leader-follower interaction can be effectively
captured via Stackelberg gradient, obtained by differentiating the follower's
strategy. Experimental results on semi- and weakly-supervised classification
and named entity recognition tasks show that our model outperforms existing
approaches by large margins.Comment: NAACL 2022 (Findings
