Analyzing 5G Network Performance Using Interactive Gaming and Video Streaming Applications
The significance of the paper is to study how 5G networks respond and adapt to different application conditions.
How Much Data Do We Really Need for Question-Answering Benchmarks? A Study Based on SQuAD v2.0
This research investigates how many
question-answer pairs are required for fine-tuning
language models on question-answering (QA) tasks. We
fine-tuned nine different language models on subsets of
the SQuAD v2.0 dataset by Rajpurkar et al. (2018)
and measured the threshold at which the marginal
benefit of additional question-answer pairs diminishes.
We show that most fine-tuned language models for QA
often do not require the full SQuAD v2.0 dataset with
130,319 training samples to perform well; 78,191 samples
(60%) are in most cases enough to achieve near-peak
performance. Thus, smaller datasets may suffice to
fine-tune QA models. Competitive performance on
smaller datasets enables less resource-intensive training
of models, makes QA tasks more accessible without
requiring powerful hardware, and helps curators of
datasets to decide how much data to collect for a given
task.
Identification of the Most Frequently Asked Questions in Financial Analyst Reports to Automate Equity Research Using Llama 3 and GPT-4
This is the first systematic analysis of the structure of equity research reports (ERRs). The paper uncovers the automation potential of ERRs, and validates their automation potential with the state-of-the-art language models GPT-4 and Llama-3-70B.
The study examines 72 ERRs, categorizing 4,964 sentences into 169 unique question archetypes across five main categories: Financials, Company, Product, Stock, and Market. The researchers classify each question based on its potential for automation, distinguishing between text-extractable, database-extractable, and non-extractable information.
Key findings include:
– 78.7% of the questions in ERRs are potentially automatable.
– 48.2% of questions are text-extractable (suited for processing by large language models).
– 30.5% of questions are database-extractable.
– Only 21.3% of questions require human judgment to answer.
The study validates these findings using two large language models: Llama-3-70B and GPT-4-turbo-2024-04-09. The results show that these models can extract relevant information from annual reports for a significant portion of the questions, with potential for even higher accuracy when used in combination.
An Interpretable Transformer-Based Approach to Classify Malaria From Blood Cell Images
Malaria is a disease that can be fatal, and it is spread through the bite of the female Anopheles mosquito. The life of the sufferer is put in jeopardy as a result of the presence of numerous plasmodium parasites, which spread throughout their blood cells. Malaria can potentially be fatal if it is not treated within the first few stages of the disease. A well-known method for diagnosing malaria, microscopy involves taking blood samples from the patient, calculating the number of parasites, and counting the victim’s red blood cells. Nevertheless, the procedure of microscopy takes a lot of time, and, in certain circumstances, it can produce an incorrect result. When compared to the more conventional approach of microscopic examination, the recent successes of deep learning (DL) in the field of medical diagnosis make it quite conceivable to reduce the expenses associated with the diagnosis while simultaneously improving overall detection accuracy. This study proposes a transformer-based DL technique for diagnosing the malaria parasite using blood cell images. An explainable AI technique called Grad-CAM was applied in order to determine which aspects of an image the proposed model paid significantly more attention to in comparison to the other aspects of the image through saliency mapping. This was done in order to demonstrate the usefulness of the models. According to the findings of this research, the performance of the vision transformer and the VGG16 are identical. Both models have reached an accuracy score of approximately 96%, which is very impressive.
ResiPlant-5: A CNN model for disease detection in citrus fruits and leaves
This paper introduces ResiPlant-5, a 5-layer con-
volutional neural network (CNN) designed for precise plant
disease diagnosis. Sequential CNN model suffer form vanishing
gradient problem. To overcome vanishing gradients and im-
prove deep model learning, the design uses skip connections,
inspired by Residual Networks (ResNets). Skip connections
provide connections between non-adjacent layers, improving
gradient propagation and feature retention. This approach lets
the model maintain important properties from previous layers
while training deeper networks without sacrificing speed. Using
deep learning and residual connections, ResiPlant-5 successfully
addresses difficult image classification challenges, hence making
it feasible to identify the disease in the plants. The model has been
trained and tested using two publicly available datasets. The first
dataset is the citrus dataset, which contains images of citrus leaves
and citrus fruit. The second dataset is the sweet orange dataset.
The results indicate that the proposed model demonstrates an
approximate increase in accuracy of 2%, 6%, and 8% on the
Sweet Orange, citrus leaves, and citrus fruit datasets, respectively,
compared to the VGG16, VGG19, and ResNet50 models
Power System Faults Analysis, Detection, and Localization in Underground Distribution and Transmission Networks by Deploying AI-based Matlab Model/Simulink
The study integrates advanced AI algorithms into MATLAB/Simulink to enhance the accuracy and efficiency of fault detection and localization. It specifically addresses the unique challenges of underground distribution and transmission networks, which are often harder to monitor and diagnose than overhead systems. The research utilizes MATLAB/Simulink for detailed simulations, enabling robust testing and validation of fault detection and localization methods under various conditions.
PMU-Based Short Circuit Capacity Estimation using System and Load Impedance Variation Ratio
This paper presents a new method for the estimation of short-circuit capacity using PMU (Phasor Measurement Units) measurements of voltage and current phasors. The proposed method has two functions: the improvement of the estimation accuracy by selecting only PMU measurements which are suitable for the estimation, and the determination of whether the estimation values are erroneous or not. The proposed method focuses on variations in system side Thevenin equivalent (TE) impedance and load impedance, which affect the estimation accuracy. First, it analyses how the changes in each component of TE or load impedance affect the short-circuit estimation. Based on these results, the proposed method is developed. Then, the proposed method is validated by numerical simulations, and it is confirmed that the proposed method can properly distinguish between the correct and incorrect estimates and obtain more accurate estimation values of short-circuit capacity than other existing methods.
Detection of wood cross-section regions with GrowCut for measurement of wood diameter grade
Proposal of a Method for Detecting the Cross-sectional Region and Measuring the Minimum Diameter of a Single Timber Cross-section Image and Verification of the Detection Accuracy
Improvement in detection accuracy was confirmed by modifying the confidence level in GrowCut.
The effectiveness of forward-seeded pixels in reducing the influence of false positives in seeded pixels was confirmed.
The effectiveness of the RANSAC method for removing false detection contour points was confirmed.
Elevating 5G Applications Performance: Harnessing Beamforming with MIMO Antennas
This paper centers on the significance of using beamforming over Multiple Input Multiple Output (MIMO) at the physical layer level within a Radio Access Network (RAN) for 5G network performance.
Semantic Segmentation of Food through Deep Learning: A Case Study
In this paper, we explore food image semantic segmentation at different levels. We identify two datasets that can be used for training and evaluating models. We also assess the performance of four semantic segmentation models for food segmentation tasks. Our results show that FCN and SegFormer achieve the best overall accuracy at 85.9% and 84.1% when applied to the UECFOODPIXCOMPLETE dataset. The study aims to offer valuable insights and guide future developments in dietary assessment tools that can underpin health management applications.