| 比賀江 文子 | D, 中間発表 | 数理情報学 | 池田 和司, | 笠原 正治, | 久保 孝富, | 日永田 智絵 | |
| TIU BENEDICT RYAN CHUA | D, 中間発表 | 数理情報学 | 池田 和司, | 笠原 正治, | 久保 孝富, | 日永田 智絵, | Li Yuzhe |
|
title: Improving macroeconomic monitoring and prediction using
Twitter data and hybrid DFM-ML models
abstract: Economic nowcasting, pertaining to forecasting the present or near future, is a common part of any economic planner's toolkit. It is especially useful during times of distress as it unfolds to ascertain the state of the economy, so as to prescribe appropriate measures to address it. However, official statistical data used as inputs often come with significant data release lags, jeopardizing the usefulness of the nowcasts. The recent COVID-19 crisis brought this problem to the fore, where near real-time proxies of the state of the economy would have helped to measure both the direction and magnitude of economic growth, as in the speed of contraction or recovery. Text data, in the form of tweets, provide a real-time stream of unstructured data that can be harnessed as alternative inputs especially in developing countries where reliable and timely economic data is scarce. Furthermore, the standard nowcasting models are limited to Factor or VAR-based models in order to deal with the varying availability of data, which has prevented the use of machine learning methods. This research thus constructs novel tweet-based indicators of economy and uncertainty, and proposes methodological advances to the modeling by using regularization in the Dynamic Factor Model (DFM), as well as combining the DFM with machine learning models. The results show smaller out-of-sample prediction errors when using the proposed models compared to the benchmark DFM model, showing promise for actual application. language of the presentation: English | |||||||
| THUWIBA QAIS ABDALLA IBRAHIM | D, 中間発表 | 数理情報学 | 池田 和司, | 荒牧 英治, | 久保 孝富, | 日永田 智絵, | Li Yuzhe |
| ISLAM MD SIHABUL | M, 2回目発表 | ディペンダブルシステム学 | 井上 美智子, | 中島 康彦, | 江口 僚太 |
|
title: Enhancing the Reliability of Memristor Crossbar-Based Neuromorphic Computing System
abstract: Neural networks (NNs) are integral to a wide range of modern computational tasks, providing advanced capabilities in areas such as pattern recognition, data analysis, and decision-making. To accelerate the computational task of NNs, memristor-based crossbar (MBC) has emerged as a promising device for accelerating multiply-accumulate (MAC) operations. However, the presence of stuck-at-faults (SAFs) in MBCs poses a significant challenge, as these faults can substantially degrade the inference accuracy of neuromorphic computing systems (NCS). We investigate some challenges of the existing SAF-tolerance method such as re-training and mapping. Although the re-training method tolerates SAFs, it is a time-consuming process to train for different faulty MBCs. Moreover, several mapping methods use redundant MBCs, and some methods tackle SAFs by minimizing the sensitivity variations of the NN model’s weight without using redundant MBCs. In our research, we propose a reliability-aware design framework for an NCS without re-training. The framework has consisted of SAF-injected training and efficient mapping without redundant hardware. To assess the framework, we conduct several experiments using an MLP and the AlexNet model. According to the experimental results, the proposed framework increases the reliability of NCS by acquiring high inference accuracy from the models for several distinct faulty MBC devices. language of the presentation: English | |||||