Temperature forecasting has been a consistent research topic owing to its significant effect on daily lives and various industries. However, it is an ever-challenging task because temperature is affected by various climate factors. Research on temperature forecasting has taken one of two directions: time-series analysis and machine learning algorithms. Recently, a large amount of high-frequent climate data have been well-stored and become available. In this study, we apply three types of neural networks, multilayer perceptron, recurrent, and convolutional, to daily average, minimum, and maximum temperature forecasting with higher-frequency input features than researchers used in previous studies. Applying these neural networks to the observed data from three locations with different climate characteristics, we show that prediction performance with highly frequent hourly input data is better than forecasting performance with less-frequent daily inputs. We observe that a convolutional neural network, which has been mostly employed for processing satellite images rather than numeric weather data for temperature forecasting, outperforms the other models. In addition, we combine state of the art weather forecasting techniques with the convolutional neural network and evaluate their effects on the temperature forecasting performances.

This paper presents an always-on Complementary Metal Oxide Semiconductor (CMOS) image sensor (CIS) using an analog convolutional neural network for image classification in mobile applications. To reduce the power consumption as well as the overall processing time, we propose analog convolution circuits for computing convolution, max-pooling, and correlated double sampling operations without operational transconductance amplifiers. In addition, we used the voltage-mode MAX circuit for max pooling in the analog domain. After the analog convolution processing, the image data were reduced by 99.58% and were converted to digital with a 4-bit single-slope analog-to-digital converter. After the conversion, images were classified by the fully connected processor, which is traditionally performed in the digital domain. The measurement results show that we achieved an 89.33% image classification accuracy. The prototype CIS was fabricated in a 0.11 μm 1-poly 4-metal CIS process with a standard 4T-active pixel sensor. The image resolution was 160 × 120, and the total power consumption of the proposed CIS was 1.12 mW with a 3.3 V supply voltage and a maximum frame rate of 120.

In a car accident, negligence is evaluated through a process known as split liability assessment. This assessment involves reconstructing the accident scenario based on information gathered from sources such as dashcam footage. The final determination of negligence is made by simulating the information contained in the video. Therefore, accident cases for split liability assessment should be classified based on information affecting the negligence degree. While deep learning has recently been in the spotlight for video recognition using short video clips, no research has been conducted to extract meaningful information from long videos, which are necessary for split liability assessment. To address this issue, we propose a new task for analysing long videos by stacking the important information predicted through the 3D CNNs model. We demonstrate the feasibility of our approach by proposing a split liability assessment method using dashcam footage.

Background: Artificial intelligence–augmented electrocardiogram (AI-ECG) models for detecting left ventricular systolic dysfunction (LVSD) often exhibit degraded performance in patients with comorbidities. Objective: This study aimed to introduce and validate a recalibration method using longitudinal patient data to enhance prediction accuracy and simulate its clinical utility for ongoing monitoring. Methods: We conducted a multicenter, retrospective cohort study using data from 2 hospitals in Korea. A dataset of paired transthoracic echocardiograms (TTEs) and electrocardiograms (ECGs) matched within a 2-week interval was constructed, separating pairs into baseline (first for each patient) and follow-up assessments. In addition to conventional supervised learning, we developed a patient-wise recalibration strategy that incorporated historical left ventricular ejection fraction measurements and prior AI-ECG outputs to adjust for future predictions, thus empirically mitigating confounding effects. Pretraining was also implemented to enhance the model’s performance. Results: The recalibrated 12-lead DeepECG LVSD model achieved an area under the receiver operating curve of 0.956 (95% CI 0.946‐0.965) for internal validation and 0.940 (95% CI 0.936‐0.945) for external validation of follow-up TTE-ECG pairs. The uncalibrated 12-lead DeepECG LVSD model also showed modest performance, with an area under the receiver operating curve of 0.953 (95% CI 0.941‐0.965) in the internal validation and 0.947 (95% CI 0.943‐0.951) in the external validation when tested on baseline TTE-ECG pairs. Recalibration yielded statistically significant improvements in the 12-lead DeepECG LVSD models ( P <.001), with enhanced and more balanced performance across all clinical subgroups. Conclusions: Patient-wise recalibration improved accuracy and consistency across various comorbidities by mitigating performance degradation and bias. This broadens the application of AI-ECG for LVSD detection from low-risk screening to high-risk longitudinal monitoring.

Multitask learning is a methodology to boost generalization performance and also reduce computational intensity and memory usage. However, learning multiple tasks simultaneously can be more difficult than learning a single task because it can cause imbalance among tasks. To address the imbalance problem, we propose an algorithm to balance between tasks at the gradient level by applying gradient-based meta-learning to multitask learning. The proposed method trains shared layers and task-specific layers separately so that the two layers with different roles in a multitask network can be fitted to their own purposes. In particular, the shared layer that contains informative knowledge shared among tasks is trained by employing single gradient step update and inner/outer loop training to mitigate the imbalance problem at the gradient level. We apply the proposed method to various multitask computer vision problems and achieve state-of-the-art performance.