136879 research outputs found
Sort by
Are AI-Generated Texts Detectable? An Experimental Study using the LibAUC Library
Get PDFAs large language models (LLMs) continue to grow and improve, it becomes harder to distinguish between AI and human written texts. The growing popularity of LLMs has also led to numerous models being developed, such as OpenAI's ChatGPT or Meta's Llama models. Identifying LLM-generated texts has practical applications across several fields, such as academia, industry, and law. However, there are many challenges, such as adversarial texts, imbalanced datasets, and out-of-distribution (OOD) performance (i.e., generalizability of detection models). Existing detectors were trained on balanced datasets and optimized loss functions that do not account for imbalanced datasets. The LibAUC library, a deep learning library for X-risk optimization, also performs strongly on imbalanced datasets. We applied the LibAUC library to train transformer models for text classification on two LLM/human text datasets. The first is the AuTexTification detection subtask (English only) dataset, which covers three domains: Wikipedia-style paragraphs, tweets, and legal documents. We evaluate model performance on a test set spanning two external domains: news and reviews. Despite this OOD test set, we show that our LibAUC-trained DeBERTaV3-large model can achieve a higher macro-F1 score than the current top scorer on the leaderboard. We also demonstrate a LibAUC-trained transformer model's performance on adversarial texts, specifically on student essays in the OUTFOX dataset. To demonstrate the scalability of our approach, we train a DistilRoBERTa model on the large-scale Deepfake Text dataset. Finally, we developed our adversarial dataset through Amazon Web Services Bedrock, generating essays from eight LLMs. We evaluated performance on this AWS-generated dataset. Our results demonstrate that essays generated by the AWS Titan models are among the most difficult to detect for our classifiers, while Claude and Llama 2's texts are easier to detect
Fertility Rates in South Asia and South Asian Diaspora in the United States: Convergence or Divergence?
No full textPakistan, India, and Bangladesh are three of the most populous South Asian countries. They were all one country till the partition of India and Pakistan in 1947 and then the partition of Pakistan and Bangladesh in 1971. Despite sharing a common history and cultural background, the three countries have followed fairly different and distinct fertility transitions. This three-manuscript dissertation focuses on fertility differentials between India, Pakistan, and Bangladesh and how these differentials might be sustained in the diaspora population within the United States and how fertility levels might differ between the Pakistani, Indian and Bangladeshi immigrant population in the United States and population residing in their home countries. The first manuscript uses Poisson regression, Oaxaca-Blinder Decomposition and Path modeling to analyze the change in fertility trends over time for each of the three countries and between countries, and the impact of the socioeconomic factors especially female education on fertility differential. Following that, I compare the recent fertility rates of Pakistani, Indian and Bangladeshi diaspora groups in the United States with other racial/ethnic groups residing in the United States. This paper uses the American Community Survey (ACS) to analyze how the process of assimilation via the increased duration of stay in the United States, improvements in the educational attainment and ability to speak English may have an impact on altering the fertility behavior of the South Asian immigrants. The final manuscript is a comparative study of recent fertility levels of Pakistani, Indian and Bangladeshi immigrants residing in the United States and their counterparts living in their home countries. This study uses ACS data to study the fertility patterns of immigrant groups and Demographic and Health Survey to study the fertility patterns of the population groups living in Pakistan, India and Bangladesh
Hemispheric Influence on Learning and Consolidation of a Dynamic Pattern With 90 Degree Relative Phase
Get PDFStudying the underlying mechanisms of bimanual coordination can extend the understanding of how we learn and control motor actions. Accounting for the integration of motor representations and sensory information from each hand, encoding the memory of a bimanual skill is often more complicated and involves more interhemispheric interaction than a unimanual skill. To date, the underlying mechanism of bimanual memory formation remains unclear. Therefore, the current dissertation was designed to address this issue. In Experiment 1, we used an interlimb transfer paradigm to study whether a 90^o relative phase (RP) pattern acquired in a unimanual condition transfers to an untrained bimanual condition. Using the same experimental design, Experiment 2 examined training-induced changes in cortical excitability (CE) at the primary motor cortex (M1) for each hemisphere using transcranial magnetic stimulation (TMS). In Experiment 3, we examined interhemispheric inhibition (IHI) between the left and right M1 after training. The results of Experiment 1 indicated that the ability to produce the 90^o RP coordinated pattern was improved across practice trials. It appeared the abstract motor representation of 90^o RP generalized to bimanual performance. In Experiment 2, we found increased CE in the left and right M1 after bimanual training, while increased CE was lateralized to the left M1 after unimanual training. This asymmetric change in CE after unimanual training indicates that the left M1 may play a critical role in forming motor memory. In Experiment 3, the results of IHI indicated an asymmetric change in IHI between each hemisphere after training. The results also indicated changes in IHI after unimanual training diminished the inhibitory influence from the right to the left M1, while IHI remained unchanged after bimanual training. The findings of the current dissertation were consistent with previous research indicating the importance of the left hemisphere in controlling unimanual and bimanual movements
Energy Efficiency/Renewable Energy Impact in the Texas Emissions Reduction Plan (TERP), Volume II ��� Technical Appendix, Annual Report to the Texas Commission on Environmental Quality January 2023 ��� December 2023
Get PDFThe Energy Systems Laboratory (ESL) at the Texas A&M Engineering Experiment Station of the Texas A&M University System is pleased to provide its annual report, ���Energy Efficiency/Renewable Energy Impact in the Texas Emissions Reduction Plan (TERP),��� as required under Texas Health and Safety Code 386.205, 386.252, 388.006, 389.003 (e), and under Texas Utilities Code Sec. 39.9051 (g) (h), and Sec. 39.9052 (c) (d)
Machine Learning Techniques for Early Identification of Timing Critical Flip-Flops in Digital IC Designs
Get PDFThe timing criticality of flip-flops is a key factor for combinational circuit timing optimization and clock network power reduction, both of which are often performed prior to CTS (Clock Tree Synthesis) and routing. However, timing criticality is often changed by CTS/routing and therefore optimizations according to pre-CTS criticality may deviate from the correct directions.
This work investigates machine learning techniques for pre-CTS identification of post-routing timing critical flip-flops. The training data will be extracted from vendor tools used for the synthesis flow and different types of machine learning engines will be evaluated and compared.
Results show that the ML-based early identification can achieve 99.7% accuracy and 0.98 area under ROC (Receiver Operating Characteristic) curve, and is 62000�� to 73000�� faster than the estimate with CTS and routing flow on average
Friction Stir Processing of WE43 with In-situ Cooling
Get PDFIt is well known that friction stir processing refines the microstructure of metallic alloys by imparting severe plastic deformation under frictional heating. FSP routinely produces an ultra-fine grained stir zone, which leads to higher hardness and yield strength in a material, and overlapping passes of FSP can be deployed to process a large area of a workpiece. It is also well known that controlling excessive heat input to the material during friction stir processing can help reduce dynamic grain growth, which is detrimental to the resulting mechanical properties. This thesis aimed to enhance the extent of grain refinement achievable during friction stir processing of a vital magnesium alloy, i.e., WE43, by leveraging active cooling. In the present work, we compared the conventional FSP with cFSP and found that fine and ultra-fine grains of 3.3 ��m and 1.5 ��m were produced in the WE43 plate by the two processes. We elaborate on the effectiveness of incorporating a cooling system into the backing plate, which provides in-situ cooling during FSP (cFSP) and thus arrests the microstructure and can help avoid grain growth during processing. cFSP grains exhibit 104 HV hardness versus the 78 HV of the base metal and improved the ductility from 9.2% Elongation of BM to 33% Elongation. The results establish that cFSP has the potential to create high hardness and ductility without compromising the yield strength which opens avenues for future research and potential applications
A Multiscale, Multifaceted Approach to Understanding Water Resources Planning and Management
No full textManaging water to provide for human and environmental needs is of paramount importance requiring the integration of biophysical, environmental, and social sciences. Further challenges to water management are presented by the uncertainties of climate change, population growth, urbanization, loss of agricultural land and desertification. These challenges exist at international, national, regional, state, and local levels. This dissertation explores three segments of water resources planning at regional, state, and local scales to better understand how planning is, and may be used to solve water management conflicts and achieve goals. Given the uncertainty of the magnitude with which climate change will impact our water resources it is important to understand how those already experiencing its effects plan for the future. An in-depth evaluation of climate change considerations in western state water plans is performed and ranking reveals that some states are taking greater precautions than others. At the state level citizen participation in, and support for financing water resources projects is assessed through Texas��� water funding referendum process which asks citizens to approve the sale of general obligation bonds which are then used to finance water projects in the state. Despite the dominance of fiscally conservative Republicans in the state, who would be expected to defeat these amendments, all seven amendments evaluated were ultimately passed thus demonstrating an understanding of the importance of financing water projects. Lastly at the local scale a demand management strategy focused on outdoor water conservation in public landscapes is introduced. As water governance shifts to small-scale projects and policies, local involvement becomes even more important. By determining the conservation potential of parks and recreational spaces and calling attention to good and bad practices cities can save water while also encouraging citizens to do the same. Water is essential and must be managed in such a manner that recognizes this fact. Achieving society���s overall water management goals is going to require understanding of and participation in water resources planning