This demo presents the next-generation version of KimBilet.com, an educational platform that leverages generative AI to create adaptive, mastery-driven learning experiences. Building on last year's personalized course generation, the new system introduces a fine-grained skill taxonomy and a feedback loop that evaluates and responds to learner performance in real time. The adaptive algorithm begins with a user's prompt to generate a topic and a taxonomy of skills needed to master that topic. For each skill, the system creates a focused lesson and a quiz. At the end of the course, learners receive a mastery profile displaying their performance across all skills. Based on these results, the system dynamically adapts: skills mastered with high accuracy are marked complete, while weaker areas trigger the generation of additional remedial lessons and quizzes until mastery is achieved. This feedback loop transforms static AI-generated lessons into personalized, mastery-based learning pathways. The demo will showcase how KimBilet.com combines generative AI, adaptive skill assessment, and iterative content refinement to engage learners across domains such as standardized test preparation, academic coursework, and professional upskilling. Participants will explore how this approach supports scalable, individualized education and discuss its potential applications in diverse learning environments.

NAIP is a collaborative project between social scientists and computer scientists to facilitate and promote datadriven research on nonprofit organizations. A major thrust of NAIP is the treatment of socioeconomic datasets such as IRS from 990 and census data for statistical analysis on nonprofits. This paper describes recent advances and experiences using Large Language Models (LLMs) to facilitate data cleaning, modeling, and automated SQL query generation on NAIP datasets. With non-trivial prompt engineering, LLMs become capable of generating useful SQL queries, but significant efforts are still needed for evaluation and validation.

This demo presents KimBilet.com, an educational platform that utilizes generative AI to create personalized educational content on demand. Catering to high-school and college students, instructors, job seekers, and lifelong learners, the system generates customized courses based on user prompts, covering any topic of interest. Each course may include a sequence of AI-created lessons and quizzes, providing detailed feedback for every quiz option to enhance understanding. The platform supports intuitive navigation through keyboard shortcuts and allows users to jump between course items seamlessly. It also maintains a history of completed quizzes to help users track their learning progress. Future enhancements include topic suggestions based on past interests, support for coding exercises, and multilingual support. This demo will showcase how KimBilet.com leverages AI to offer adaptive learning experiences, engage attendees through interactive exploration, and discuss its potential applications in educational settings. Participants will gain insights into integrating AI-driven tools into teaching and learning processes to address diverse educational needs.

Nonprofit organizations are important contributors to the US economy and social well-being. The Nonprofit Organization Research Panel Project (NORPP) Manager has been developing and sharing datasets and software tools to facilitate data-driven research on nonprofit organizations. The project has two major thrusts: (1) large-scale survey panels, e.g., the Annual National Survey of Nonprofit Trends and Impacts, from 2021 to the current (2024); and (2) NORPP Analytics Integration Platform (NAIP) to collect, process, and query a wide range of socioeconomic indicator datasets, such as IRS 990 form and census data. Technical challenges that include data heterogeneity, data quality, and protection of sensitive data have made the expansion and maintenance of NAIP datasets both labor-intensive and time-consuming. We are currently exploring new technologies, including Large Language Models such as GPT series, to automate database query generation, schema adaptation, and data quality assurance processes.

Fine-tuning pre-trained large language models (LLMs) has become a common practice for personalized natural language understanding (NLU) applications on downstream tasks and domain-specific datasets. However, there are two main challenges: (i) limited and/or heterogeneous data for fine-tuning due to proprietary data confidentiality or privacy requirements, and (ii) varying computation resources available across participating clients such as edge devices. This paper presents FedHFT - an efficient and personalized federated fine-tuning framework to address both challenges. First, we introduce a mixture of masked adapters to handle resource heterogeneity across participating clients, enabling high-performance collaborative fine-tuning of pre-trained language model(s) across multiple clients in a distributed setting, while keeping proprietary data local. Second, we introduce a bi-level optimization approach to handle non-iid data distribution based on masked personalization and client clustering. Extensive experiments demonstrate significant performance and efficiency improvements over various natural language understanding tasks under data and resource heterogeneity compared to representative heterogeneous federated learning methods.

Maintaining consistently low response times is a critical performance requirement for mission-critical, web-facing applications (e.g., e-commerce) that typically use microservices architectures. Through extensive benchmarking of a microservices application in a cloud environment, we demonstrate that its response time stability is fragile, with significant variations (ranging from milliseconds to seconds) when the CPU utilization of component servers reaches moderate to high levels (e.g., 60%). Our detailed timeline analysis identifies a leading cause of response time instability at these utilization levels: a ripple effect caused by the long chain of dependencies inherent in microservices applications. In such an architecture, a millibottleneck (a bottleneck lasting sub-seconds) can trigger a queuing effect from a downstream server that propagates to upstream servers, resulting in dropped requests and TCP retransmissions lasting several seconds at the weakest link in the chain. We demonstrate two common factors that contribute to millibottlenecks in microservices runtime environments: interference from collocated containers and bursty workloads, both of which are prevalent in real-world cloud environments.

The modern web services landscape is characterized by numerous fine-grained, loosely coupled microservices with increasingly stringent low-latency requirements. However, this architecture also brings new performance vulnerabilities. In this paper, we introduce a novel low-volume application layer DDoS attack called the Sync-Millibottleneck (SyncM) attack, specifically targeting microservices. The goal of this attack is to cause a long-tail latency problem that violates the service-level agreement (SLA) while evading state-of-the-art DDoS detection/defense mechanisms. The SyncM attack exploits two unique features of microservices architecture: (1) the shared frontend gateway that directs user requests to mid-tier/backend microservices, and (2) the co-existence of multiple logically independent execution paths, each with its own bottleneck resource. By creating synchronized millibottlenecks (i.e., sub-second duration bottlenecks) on multiple independent execution paths, SyncM attack can cause the queuing effect in each execution path to be propagated and superimposed in the shared frontend gateway. As a result, SyncM triggers surprisingly high latency spikes in the system, even when all system resources are far from saturation, making it challenging to trace the cause of performance instability. To evaluate the practicality of the SyncM attack, we conduct extensive experiments on real cloud systems such as EC2 and Azure, which are equipped with state-of-the-art IDS/IPS systems. We also conduct a large-scale simulation using a production Alibaba trace to show the scalability of our attack. Our results demonstrate that the SyncM attack is highly effective, as it only consumes less than 15% of additional CPU resources of the target system while increasing its 95th percentile response time by more than 20 times.

Deep learning shows great promise in providing more intelligence to the cross-platform web. However, insufficient infrastructure, heavy models, and intensive computation limit the use of deep learning with low-performing web browsers. We propose DeepAdapter, an integrated cloud-edge-device framework that ties the edge, the remote cloud, with the device by cross-platform web technology for adaptive deep learning services towards lower latency, lower mobile energy, and higher system throughput. DeepAdapter consists of context-aware pruning, service updating, and online scheduling. First, the offline pruning module provides a context-aware pruning algorithm that incorporates the latency, the network condition, and the device's computing capability to fit various contexts. Second, the service updating module optimizes branch model cache on the edge for massive mobile users and updates the new model pruning requirements. Third, the online scheduling module matches optimal branch models for mobile users. Also, a two-stage DRL-based online scheduling method named DeepScheduler can handle high concurrent requests between edge centers and remote cloud by designing the reward prediction model. Extensive experiments show that DeepAdapter can decrease average latency by 1.33x, reduce average mobile energy by 1.4x, and improve system throughput by 2.1x with considerable accuracy.

Toxic misinformation campaigns have caused significant societal harm, e.g., affecting elections and COVID-19 information awareness. Unfortunately, despite successes of (gold standard) retrospective studies of misinformation that confirmed their harmful effects after the fact, they arrive too late for timely intervention and reduction of such harm. By design, misinformation evades retrospective classifiers by exploiting two properties we call new-normal: (1) never-seen-before novelty that cause inescapable generalization challenges for previous classifiers, and (2) massive but short campaigns that end before they can be manually annotated for new classifier training. To tackle these challenges, we propose UFIT, which combines two techniques: semantic masking of strong signal keywords to reduce overfitting, and intra-proxy smoothness regularization of high-density regions in the latent space to improve reliability and maintain accuracy. Evaluation of UFIT on public new-normal misinformation data shows over 30% improvement over existing approaches on future (and unseen) campaigns. To the best of our knowledge, UFIT is the first successful effort to achieve such high level of generalization on new-normal misinformation data with minimal concession (1 to 5%) of accuracy compared to oracles trained with full knowledge of all campaigns.

COVID-19 related misinformation and fake news, coined an 'infodemic', has dramatically increased over the past few years. This misinformation exhibits concept drift, where the distribution of fake news changes over time, reducing effectiveness of previously trained models for fake news detection. Given a set of fake news models trained on multiple domains, we propose an adaptive decision module to select the best-fit model for a new sample. We propose MiDAS, a multi-domain adaptative approach for fake news detection that ranks relevancy of existing models to new samples. MiDAS contains 2 components: a doman-invariant encoder, and an adaptive model selector. MiDAS integrates multiple pre-trained and fine-tuned models with their training data to create a domain-invariant representation. Then, MiDAS uses local Lipschitz smoothness of the invariant embedding space to estimate each model's relevance to a new sample. Higher ranked models provide predictions, and lower ranked models abstain. We evaluate MiDAS on generalization to drifted data with 9 fake news datasets, each obtained from different domains and modalities. MiDAS achieves new state-of-the-art performance on multi-domain adaptation for out-of-distribution fake news classification.

The vehicle recognition area, including vehicle make-model recognition (VMMR), re-id, tracking, and parts-detection, has made significant progress in recent years, driven by several large-scale datasets for each task. These datasets are often non-overlapping, with different label schemas for each task: VMMR focuses on make and model, while re-id focuses on vehicle ID. It is promising to combine these datasets to take advantage of knowledge across datasets as well as increased training data; however, dataset integration is challenging due to the domain gap problem. This paper proposes ATEAM, an annotation team-of-experts to perform cross-dataset labeling and integration of disjoint annotation schemas. ATEAM uses diverse experts, each trained on datasets that contain an annotation schema, to transfer knowledge to datasets without that annotation. Using ATEAM, we integrated several common vehicle recognition datasets into a Knowledge Integrated Dataset (KID). We evaluate ATEAM and KID for vehicle recognition problems and show that our integrated dataset can help off-the-shelf models achieve excellent accuracy on VMMR and vehicle re-id with no changes to model architectures. We achieve mAP of 0.83 on VeRi, and accuracy of 0.97 on CompCars. We have released both the dataset and the ATEAM framework for public use.

Recent advances in text classification and knowledge capture in language models have relied on availability of large-scale text datasets. However, language models are trained on static snapshots of knowledge and are limited when that knowledge evolves. This is especially critical for misinformation detection, where new types of misinformation continuously appear, replacing old campaigns. We propose time-aware misinformation datasets to capture time-critical phenomena. In this paper, we first present evidence of evolving misinformation and show that incorporating even simple time-awareness significantly improves classifier accuracy. Second, we present COVID-TAD, a large-scale COVID-19 misinformation da-taset spanning 25 months. It is the first large-scale misinformation dataset that contains multiple snapshots of a datastream and is orders of magnitude bigger than related misinformation datasets. We describe the collection and labeling pro-cess, as well as preliminary experiments.

The Covid-19 pandemic has caused a dramatic and parallel rise in dangerous misinformation, denoted an ‘infodemic’ by the CDC and WHO. Misinformation tied to the Covid-19 infodemic changes continuously; this can lead to performance degradation of fine-tuned models due to concept drift. Degredation can be mitigated if models generalize well-enough to capture some cyclical aspects of drifted data. In this paper, we explore generalizability of pre-trained and fine-tuned fake news detectors across 9 fake news datasets. We show that existing models often overfit on their training dataset and have poor performance on unseen data. However, on some subsets of unseen data that overlap with training data, models have higher accuracy. Based on this observation, we also present KMeans-Proxy, a fast and effective method based on K-Means clustering for quickly identifying these overlapping subsets of unseen data. KMeans-Proxy improves generalizability on unseen fake news datasets by 0.1-0.2 f1-points across datasets. We present both our generalizability experiments as well as KMeans-Proxy to further research in tackling the fake news problem.

Multi-user mobile Augmented Reality (AR) has been successfully used in various fields as a novel visual interaction technology. But current mainstream wearable device-based and app-based solutions are still facing cross-platform, real-time communication, and intensive computing requirements. Mobile Web technology is envisioned to be a promising supporting technology for cross-platform application of mobile AR especially in 5G networks, which provide pervasive communication and computing resources thereby forming a formidable framework for the practical application of multi-user mobile Web AR. However, the problem of how to use these new techniques properly to achieve efficient communication and computing collaboration is obviously paramount in order for multi-user mobile Web AR to be realized in 5G networks. In this article, we propose the first edge-assisted multi-user collaborative framework for mobile Web AR in the 5G era. First, we propose a heuristic mechanism BA-CPP for efficient communication planning, which allows multi-user interaction synchronization to be achieved. Second, we introduce a motion-aware key frame selection mechanism called Mo-KFP to optimize the computational efficiency of the edge system, and simultaneously alleviate the initialization problem by collaborating with nearby mobile devices using the Device-to-Device (D2D) communication technique. Experiments are conducted in a real-world 5G network, and the results demonstrate the superiority of our proposed collaborative framework.

Enabling deep learning technology on the mobile web can improve the user's experience for achieving web artificial intelligence in various fields. However, heavy DNN models and limited computing resources of the mobile web are now unable to support executing computationally intensive DNNs when deploying in a cloud computing platform. With the help of promising edge computing, we propose a lightweight collaborative deep neural network for the mobile web, named LcDNN, which contributes to three aspects: (1) We design a composite collaborative DNN that reduces the model size, accelerates inference, and reduces mobile energy cost by executing a lightweight binary neural network (BNN) branch on the mobile web. (2) We provide a jointly training method for LcDNN and implement an energy-efficient inference library for executing the BNN branch on the mobile web. (3) To further promote the resource utilization of the edge cloud, we develop a DRL-based online scheduling scheme to obtain an optimal allocation for LcDNN. The experimental results show that LcDNN outperforms existing approaches for reducing the model size by about 16x to 29x. It also reduces the end-to-end latency and mobile energy cost with acceptable accuracy and improves the throughput and resource utilization of the edge cloud.

Mission-critical, real-time, continuous stream processing applications that interact with the real world have stringent latency requirements. For example, e-commerce websites like Amazon improve their marketing strategy by performing real-time advertising based on customers' behavior, and latency long tail can cause significant revenue loss. Recent work [39] showed a positive correlation between latency long tail and variance in the execution time of synchronous invocation chains (critical paths) in microservices benchmarks. This paper shows that asynchronous, very short but intense resource demands (called millibottlenecks) outside of critical paths can also cause significant latency long tail. Using a traffic analysis stream processing application benchmark, we evaluated the impact of asynchronous workload bursts generated by a multi-layer data structure called LSM-tree (log-structured merge-tree) for continuous checkpointing. Outside of the critical path, LSM-tree relies on maintenance operations (e.g., flushing/compaction during a checkpoint) to reorganize LSM-tree in memory and on disk to keep data access latency short. Although asynchronous, such recurrent maintenance operations can cause frequent millibottlenecks, particularly when they overlap, a problem we call ShadowSync. For scheduling and statistical reasons, significant latency long tail can arise from ShadowSync caused by asynchronous recurrent operations. Our experimental results show that with typical settings of benchmark components such as RocksDB, ShadowSync can prolong request message latency by up to 2 seconds. We show effective mitigation methods can alleviate both scheduled and statistical ShadowSync reducing the latency long tail to less than 20% of the original at the 99.9 th percentile.

Machine Learning has become the bedrock of recent advances in text, image, video, and audio processing and generation. Most production systems deal with several models during deployment and training, each with a variety of tuned hyperparameters. Furthermore, data collection and processing aspects of ML pipelines are receiving increasing interest due to their importance in creating sustainable high-quality classifiers. We present EdnaML, a framework with a declarative API for reproducible deep learning. EdnaML provides low-level building blocks that can be composed manually, as well as a high-level pipeline orchestration API to automate data collection, data processing, classifier training, classifier deployment, and model monitoring. Our layered API allows users to manage ML pipelines at high-level component abstractions, while providing flexibility to modify any part of it through the building blocks. We present several examples of ML pipelines with EdnaML, including a large-scale fake news labeling and classification system with six sub-pipelines managed by EdnaML.

Big data powered Deep Learning (DL) and its applications have blossomed in recent years, fueled by three technological trends: a large amount of digitized data openly accessible, a growing number of DL software frameworks in open source and commercial markets, and a selection of affordable parallel computing hardware devices. However, no single DL framework, to date, dominates in terms of performance and accuracy even for baseline classification tasks on standard datasets, making the selection of a DL framework an overwhelming task. This paper takes a holistic approach to conduct empirical comparison and analysis of four representative DL frameworks with three unique contributions. First, given a selection of CPU-GPU configurations, we show that for a specific DL framework, different configurations of its hyper-parameters may have a significant impact on both performance and accuracy of DL applications. Second, to the best of our knowledge, this study is the first to identify the opportunities for improving the training time performance and the accuracy of DL frameworks by configuring parallel computing libraries and tuning individual and multiple hyper-parameters. Third, we also conduct a comparative measurement study on the resource consumption patterns of four DL frameworks and their performance and accuracy implications, including CPU and memory usage, and their correlations to varying settings of hyper-parameters under different configuration combinations of hardware, parallel computing libraries. We argue that this measurement study provides in-depth empirical comparison and analysis of four representative DL frameworks, and offers practical guidance for service providers to deploying and delivering DL as a Service (DLaaS) and for application developers and DLaaS consumers to select the right DL frameworks for the right DL workloads.

Machine learning models with explainable predictions are increasingly sought after, especially for real-world, mission-critical applications that require bias detection and risk mitigation. Inherent interpretability, where a model is designed from the ground-up for interpretability, provides intuitive insights and transparent explanations on model prediction and performance. In this paper, we present CoLabel, an approach to build interpretable models with explanations rooted in the ground truth. We demonstrate CoLabel in a vehicle feature extraction application in the context of vehicle make-model recognition (VMMR). By construction, CoLabel performs VMMR with a composite of interpretable features such as vehicle color, type, and make, all based on interpretable annotations of the ground truth labels. First, CoLabel performs corroborative integration to join multiple datasets that each have a subset of desired annotations of color, type, and make. Then, CoLabel uses decomposable branches to extract complementary features corresponding to desired annotations. Finally, CoLabel fuses them together for final predictions. During feature fusion, CoLabel harmonizes complementary branches so that VMMR features are compatible with each other and can be projected to the same semantic space for classification. With inherent interpretability, CoLabel achieves superior performance to the state-of-the-art black-box models, with accuracy of 0.98, 0.95, and 0.94 on CompCars, Cars196, and BoxCars116K, respectively. CoLabel provides intuitive explanations due to constructive interpretability, and subsequently achieves high accuracy and usability in mission-critical situations.

Web-based Deep Neural Networks (DNNs) enhance the ability of object recognition and has attracted considerable attention in mobile Web AR and other services. However, neither performing the DNN inference on mobile Web browsers locally nor offloading computations to the cloud can strike a balance between accuracy and efficiency; generally, rude methods are often accompanied by unsatisfactory accuracy. Collaborative approaches seem to fill this gap by coordinating the distributed hierarchical computing resources, especially in the 5G era, but it still faces challenges in the current solutions, such as the lack of (1) full use of 5G resources for the one point DNN computation partitioning schemes; (2) fine-grained branching mechanism; (3) efficient partitioning method; and (4) multi-objective optimization. To this end, we present the fine-grained elastic computation partitioning mechanism for distributed DNN in 5G networks. First, we elaborate two collaborative scenarios. Second, we study the DNN branching mechanism at layer granularity. Next, we propose a DNN computation partitioning algorithm based on deep reinforcement learning. Finally, we develop a mobile Web AR application as a proof of concept. The experiments were conducted in an actually deployed 5G trial network, and the results show the superiority of this collaborative approach. The common theme is, under the premise that Quality of Service (QoS) is satisfied, to balance multiple interests by orchestrating computations across heterogeneous computing platforms.