RSS 1.0SSI 2022 ePosters Competition
Recent Submissions
-
Statistical Modeling of Extreme PM2.5 Concentration in Saudi Arabia Using Deep Learning(2022-08-02) [Poster]
-
A new R package for evaluation of probabilistic forecasts based on locally scale invariant proper scoring rules(2022-08-02) [Poster]It is common to make predictions in the form of probabilistic forecasts in areas such as weather forecasting. Proper scoring rules are the standard way to evaluate these predictions. We implement a newly proposed proper scoring rule, which has some desirable properties, in an R package. We then use it to compare different fitted models with real data.
-
Fast Data Transfer to Mitigate Data Stall in Machine Learning Models(2022-08-02) [Poster]Deep Neural Network (DNN) is one of the most influential models that provides promising machine learning solutions. However, DNN training consumes much data, energy, and time. It is also resource-intensive as its data pipeline stages start from storage through to the CPU and end with the GPU. The first two stages involve fetching data from storage and implementing data pre-processing in the CPU, and they consume much time; this eventually leads to a data stall problem which describes leaving the GPU, in the third stage, waiting for data to be fetched or pre-processed in an idle state. To mitigate the data-stall problem in terms of pre-processing, we designed a distributed machine learning system that takes advantage of programmable in-network components to accelerate the processes and reduce the load on the CPU via pre-processing. In this project, an Alveo U280 Data Center Accelerator Card programmed with Verilog was used to perform the following common pre-processing transformations: One-hot, Clamp, Random Selection, Cartesian Product, and Timestamp.
-
Optimal Placement of Reconfigurable Intelligent Surfaces (RIS) for Wireless Systems(2022-08-02) [Poster]Reconfigurable Intelligent Surfaces (RIS) have been proposed to reconfigure the wireless propagation environment and enhance transmission performance. A RIS is made of passive reconfigurable elements and can vary the phase and amplitude of the incident signals to reflect them in the desired directions. RISs do not require power-hungry RF chains and power amplifiers, making them low-cost and energy-efficient solutions for coverage extensions. RIS-aided communications can therefore be used to extend the coverage range of actual wireless networks. In our work, we study the optimal placement of RISs for wireless systems, considering the existence of blockages in the terrain.
-
Customized mesoporous metal organic frameworks (MOFs) engender stable enzymatic nanoreactors(2022-08-02) [Poster]Synthesis and characterization of MIL-101(Fe) and its importance in the enzymatic field.
