In this paper, we provide an overview of a specific, but rather popular type of blockchain scaling solutions, which are usually referred to as Layer 2 blockchains. These solutions depend on the rollup scaling strategy. First, they all create batches of transactions known as rollups. Second, there are two primary approaches to rollup verifications. One approach is referred to as optimistic rollup, where it is assumed that fraudulent transactions rarely occur and should it happen, then some nodes could raise a challenge, which would lead to the resolution of the problem. The interaction between the asserter and the challenger is referred to as interactive fraud proof. The other approach is a conservative approach, where every rollup batch is associated with a validity proof. For all practical purposes, Ethereum is used as the Layer 1 blockchain because it is the first and only large-scale public blockchain that supports Turing-complete smart contracts. These Layer 2 blockchains typically deploy at least two smart contracts on the Layer 1 blockchain, one to receive rollups from the Layer 2 blockchain, and the other to verify the rollup batch. Furthermore, we provide technical details of two Layer 2 blockchains (i.e., Arbitrum and Optimism) that use the optimistic rollup mechanism, and two Layer 2 blockchains (i.e., Polygon and Starknet) that use validity rollup mechanism. Finally, we analyze the Layer 2 blockchain solutions in the framework of the blockchain trilemma theory. We show that all Layer 2 scaling solutions trade decentralization for better scalability.
Artificial intelligence (AI), particularly deep learning, has recently made significant strides in healthcare, offering promising tools for the early detection and management of heart disease, a leading cause of mortality globally. However, the accessibility and quality of such care are often limited in rural areas. Convolutional Neural Networks (CNNs) have been particularly successful in analyzing 12-lead electrocardiograms (ECGs) to detect cardiac anomalies crucial for accurate diagnosis. In this study, we present a comparative analysis of various deep learning models tailored for deployment in resource-constrained rural healthcare environments, emphasizing energy efficiency, low computational demands, and user-friendliness. We introduce a novel Residual Convolutional Gated Recurrent Network (Residual ConvGRU) and evaluate its performance against established benchmarks using 10-fold cross-validation on three widely recognized datasets for classifying different types of cardiac arrhythmias (CAs). We also conduct further model compression experiments on the top 3 models and our proposed model to explore potentials for future deployment in environments with limited computational resources such as in rural areas.
Transformers have become the state-of-the-art architectures for Natural Language Processing (NLP) by capturing the long-range dependencies with the help of attention mechanisms. As the use of Transformers is increasing in the field of time series analysis, we conducted our research on assessing the feasibility of applying secure inference to time series Transformers using CrypTen, a privacy-preserving framework for machine learning applications through Multi Party Computation (MPC). Our approach uses only the encoder part of Transformers for training on plaintext data and will conduct inference on the CrypTen. This methodology combines the predictive power of Transformers and the security power of CrypTen, and we propose a novel foundational framework for privacy-preserving time series analysis using Transformers. Despite adding private inference, we got performance results similar to those of Informer model. This work represents a significant step towards secure predictive analytics showcasing the potential of Transformers in privacy-preserving time series forecasting.
The objective of this research is to create an object-detecting Android application for ELISpot immunoassay reading. The ability to accurately, quickly, and cheaply conduct immunoassay tests will guarantee that public healthcare industries are operating and helping patients in rural areas effectively. The past reliance on purchased laboratory equipment, and the rampant cost of health services has led to the privatization of healthcare industries and the negligence of rural area healthcare facilities. Therefore, this research is intended to help spread ELISpot test capability to smaller less-funded programs by creating an accessible ELISpot reading system in the form of an application and the hardware required. Object detection solutions present 2 issues; computational intensity and image quality. An Android application containing the enumeration algorithm was also designed. In order to implement object detection capability on mobile devices, the parameters used were extracted from a Faster R-CNN neural network model trained from a set of immunoassay images provided by the laboratory. The annotation of the images was done using LabelImage, whose output of .xml was converted to .csv and tfrecord files respectively. The computational accuracy tested over the samples averages out at 98% accuracy. While other design parameters such as image processing speed, system weight, and process throughput time are all satisfactory. With the project completed successfully, it is believed that with the upcoming technologies of deep learning super sampling (DLSS) With this kept in mind, true deep learning capable object detection applications operated on a mobile device will soon become more widespread and accessible.
Online Action Detection aims to identify ongoing incomplete actions from online video streams. The key is to model the inter-frame relationships related to the current action, i.e., temporal dependency. However, frames with similar appearances may contain different action semantics. Therefore, modeling the attention interaction between frame features without considering the action semantics may result in incorrect dependency between frames with similar appearances but different semantics. To address this problem, we propose to utilize action semantics and frame features to jointly model temporal dependency. Specifically, to mine frame-level action semantics, our proposed Semantic Correlation Transformer (SecTR) first extracts representative features from redundant and diverse action frames features by performing K-means clustering for each action category. Considering that a single frame only contains local action information, SecTR reconstructs the frame feature based on clustering features from the perspective of various categories, which is used to mine the action semantics of frames. Then, according to consistency with the semantics of the current frame, SecTR further embeds the frame semantics into the frame features to jointly model temporal dependency. Moreover, to further optimize the modeling quality, we propose an auxiliary task called current action anticipation, which predicts the current action based on historical frames. The effectiveness and superiority of SecTR are validated on THUMOS'14, TVSeries and HDD.
