Weekly Bulletin

Change-point analysis is thriving in this big data era, addressing problems that arise across many fields where massive data sequences are collected to study complex phenomena over time. It plays a crucial role in processing these data by segmenting long sequences into homogeneous parts for subsequent studies. Observations could be high-dimensional or not lie in the Euclidean space, such as network data, which are challenging to characterize using parametric models. We utilize the inter-point information of the observations and propose a series of nonparametric methods to address the issue. In particular, we take into account a pattern caused by the curse of dimensionality so that the proposed methods can accommodate a broad range of alternatives. Additionally, we work out ways to analytically approximate the p-values of the test statistics, enabling rapid type I error control. The methods are applied to Neuropixels data in the analysis of thousands of neurons’ activities.

When we teach statistics and machine learning, we typically imagine problems in which we wish to predict some target Y from data X, or to build understanding of the relationship between these two variables, or to test some predicted effect of intervening between them. We fit models based on samples of these pairs. Yet we rarely investigate precisely where our labeled data comes from, referring instead to labels (Y) in supervised learning problems as "gold-standard" feedback, or something similar. Yet these labels are constructed via sophisticated pipelines, aggregating expert (or non-expert) feedback, combining observations in sophisticated ways, and we do not model these choices in our statistical learning pipelines. In this talk, I will discuss some work we have been doing to try to open up this bigger picture of statistics, providing some food for thought about how we might move beyond our standard statistical analyses.