Event(s) on February 2022

• Tuesday, 15th February, 2022

  Title:	Multiple Change Point Detection for High Dimensional Data
  Speaker:	Mr ZHAO Wenbiao, Department of Mathematics, Hong Kong Baptist University, Hong Kong
  Time/Place:	14:30  -  16:00 Zoom, Meeting ID: 933 5327 8327 Password: 492651
  Abstract:	In this research, we investigate simultaneously detecting multiple change points for high-dimensional data that the dimension can be of exponential rate of the sample size. The proposed estimation approach utilizes a signal statistic that is based on a sequence of local U-statistics no matter whether the data are with sparse or dense structure. It can avoid both expensive computations that exhaustive search algorithms need, and false positives that hypothesis testing-based approaches have to control. The estimation consistency can hold for the locations and number of change points even when the number of change points diverges at a certain rate as the sample size goes to infinity. Further, because of its visualization nature, in practice, plotting the signal statistic can greatly help identify the locations in contrast to existing methods in the literature. The numerical studies are conducted to examine its performance in finite sample scenarios and a real data example is analyzed for illustration.

  
  

• Thursday, 17th February, 2022

  Title:	Error Compensated Stochastic Gradient Method for Communication-Efficient Federated Learning
  Speaker:	Mr NI Renyuan, Department of Mathematics, Hong Kong Baptist University, Hong Kong
  Time/Place:	10:30  -  12:00 Zoom, Meeting ID: 929 4302 8760 Password: 925318
  Abstract:	We are considering solving a convex distributed stochastic optimization that is widely used in the field of federated learning. In this lecture, we propose a FEDPEQ method that combines three techniques: quantization, partial participation and error feedback together to reduce the communication cost in the training process. The theoretical results show that the convergence rate of our method is closely related to the number of clients used in each iteration. Finally, we implemented our algorithm to solve the image classification on several datasets.

All years