Skip to content

Latest commit

 

History

History
102 lines (99 loc) · 6.8 KB

README.org

File metadata and controls

102 lines (99 loc) · 6.8 KB
Raw

Group meeting schedule, Spring 2021

Mondays 3:30-4:30 on zoom.

  1. your research.
  2. somebody else’s paper.
  3. some useful software.
  • Jan 11: Toby, How I organize my research projects.
  • Jan 18: Holiday (no meeting).
  • Jan 25: Tristan, PeakLearner Slides, Code
  • Feb 1: Jon
  • Feb 8: David Venuto, Thinking Realistically with Deep Reinforcement Learning.
    • Abstract: Reinforcement learning (RL) gives us a framework for learning how to take intelligent actions in an environment represented as a Markov Decision Process. Recent advances have been able to learn in highly complex environments by learning policies and value functions approximated with deep neural networks. These advances come with a wide set of challenges including learning generalizable policies and designing good reward functions. We will first give a brief introduction to the paradigm of RL and how we have leveraged deep learning to improve policy optimization. Next, we will discuss common issues in deep RL and how we can learn with more robustness by leveraging expert data and learning modular and hierarchical policies.
  • Feb 15: Alyssa: FLOPART slides
  • Feb 22: Akhila, RcppDeepState Slides, Code, Blog
  • Mar 1: Vincent Runge, gfpop: an R Package for multiple change-point detection constrained by a graph.
    • Abstract: The accurate detection of multiple change-points in univariate time series is a daily task for many engineers and scientists. Often, practitioners can have prior knowledge about the type of changes they are looking for. For example in genomic data, biologists expect peaks: up changes followed by down changes. Integrating such priors is important and requires dedicated algorithms. We propose with the gfpop R package a generic algorithm able to deal with many priors constraining the successive segment means. This algorithm can be seen as a Hidden Markov Chain model with continuous state space. gfpop works for a user-defined graph of constraints and several loss functions: Gauss, Poisson, Binomial, Biweight and Huber. This presentation starts with a long introduction about multiple change-point detection by penalized dynamic programming. The gfpop penalized optimization problem is solved by the functional pruning optimal partitioning (fpop) algorithm. This dynamic programming approach returns the exact minimizer. The approximate binary segmentation method can not be used with constraints on successive segments. We illustrate the use of gfpop on isotonic simulations and several applications in biology.
  • Mar 8: Tristan: PeakLearner Labeling Activity (Fail)
  • Mar 15: Arnaud Liehrmann, Increased peak detection accuracy in over-dispersed ChIP-seq data with supervised segmentation models.
    • Abstract: Background: Histone modification constitutes a basic mechanism for the genetic regulation of gene expression. In early 2000s, a powerful technique has emerged that couples chromatin immunoprecipitation with high-throughput sequencing (ChIP-seq). This technique provides a direct survey of the DNA regions associated with these modifications. In order to realize the full potential of this technique, increasingly sophisticated statistical algorithms have been developed or adapted to analyze the massive amount of data it generates. Many of these algorithms were built around natural assumptions such as the Poisson distribution to model the noise in the count data. In this work, we start from these natural assumptions and show that it is possible to improve upon them. Results: Our comparisons on seven reference datasets of histone modifications (H3K36me3 \& H3K4me3) suggest that natural assumptions are not always realistic under application conditions. We show that the unconstrained multiple changepoint detection model with alternative noise assumptions and supervised learning of the penalty parameter reduces the over-dispersion exhibited by count data. This model detects peaks more accurately than algorithms that rely on natural assumptions. Conclusion: The segmentation model we propose can benefit researchers in the field of epigenetics by providing new high-quality peak prediction tracks for H3K36me3 and H3k4me3 histone modifications.
  • Mar 22: Alex Drouin, Differentiable Causal Discovery from Interventional Data, arXiv pre-print, video.
    • Abstract: Knowledge of the causal structure that underlies a data generating process is essential to answering questions of causal nature. Such questions are abundant in fields that involve decision making such as econometrics, epidemiology, and social sciences. When causal knowledge is unavailable, one can resort to causal discovery algorithms, which attempt to recover causal relationships from data. This talk will present a new algorithm for this task, that combines continuous-constrained optimization with the flexible density estimation capabilities of normalizing flows. In contrast with previous work in this direction, our method combines observational and interventional data to improve identification of the causal graph. We will present empirical results, along with a theoretical justification of our algorithm.
    • Short bio: Alexandre Drouin is a Research Scientist at Element AI in Montréal, Canada and an Adjunct Professor of computer science at Laval University. He received a PhD in machine learning from Laval University in 2019 for his work on antibiotic resistance prediction in bacterial genomes. His research interests include causal inference, deep learning, and bioinformatics
  • Mar 29: Frank mlr3 slides
  • Apr 5: Jon AUM Update Presentation Slides, Toby Why is functional pruning so fast for optimal changepoint detection?
  • Apr 12: Alyssa
  • Apr 19: Tristan