| 古庄 泰隆 | D, 中間発表 | 数理情報学 | 池田 和司, 松本 裕治, 吉本 潤一郎, 佐々木 博昭 |
|
title: ResNet that skips two layers slowly converges to flat minima
abstract: Neural networks have been changing the history of machine learning in performance. The high performance comes from transforming an input vector to a new feature vector through many hidden layers. Actually, approximation ability of multi layer perceptron (MLP) with ReLU activation grows polynomially with its width and exponentially with its depth. However, degradation of the empirical risk is observed when a neural network which simply stacks layers (plain-net) like MLP is too deep. In order to overcome this problem, ResNet, whose each hidden layer has skip-connection from its input to its output, was proposed. Thanks to skip-connection, extreme deep ResNet can be trained and achieved high performance. Behind this result, loss of plain-net and ResNet behaved differently regarding convergence speed of gradient descent and generalization gap. In order to explain high performance of ResNet, we analyzed effect of skip-connection on convergence speed of gradient descent and generalization gap. We show that ResNet which skips two layers slowly converges to flat minima, which achieves small generalization gap. language of the presentation: English | |||
| 井上 嵩史 | M, 2回目発表 | 数理情報学 | 池田 和司☆, 松本 裕治, 澤田 宏(客員）, 岩田 具治(客員） |
|
title: Population prediction using spatio-temporal data
abstract: In recent years, urbanization has accelerated. That makes the congestion problem more serious. Anyway, I predict the future population of specific areas in my research to solve this problem. In my presentation, I am going to report some results that are predicted by common prediction models in previous research, for example AutoRegressive (AR) model and Long Short-Term Memory (LSTM) model. In addition, I am going to show an utility of a model using location data. In conclusion, I am going to talk about a future plan of my research. language of the presentation: Japanese | |||
| 伊藤 健史 | M, 2回目発表 | 数理情報学 | 池田 和司☆, 小笠原 司, 川人 光男, 森本 淳, Nishanth Koganti |
|
title: High Frequency Feedback Control with Neural Network Dynamics Model
abstract: THIS ABSTRACT IS NOT PUBLICLY AVAILABLE DUE TO THE PROSPECTIVE PATTENT APPLICATION. language of the presentation: Japanese | |||