README.org

Group meeting schedule, Spring 2023

Time 2:30-3:30pm on Thursday afternoons in Building 90 (SICCS), room 224,

What to present?

1. your research.
2. somebody else’s paper.
3. some useful software.

• Jan 19: Toby, Title: Cross-validation for training and testing co-occurence network inference algorithms, source. Can we evaluate Chiquet’s algorithm using cross-validation? PLNmodels slides, ICML’19 paper.
• Jan 26: Bilal, Title: Vehicle Detection in Satellite Images by Hybrid Deep Convolutional Neural Networks. Discussion about how I will use this methodology for my own research, and what I have done so far in my research and way forward.
• Feb 2: Jadon, Title: Exact Line Search for AUM. I’ll talk about Area Under ROC Curves (AUC) and a differentiable surrogate loss function for AUC known as the Area Under Min(False Positives, False Negatives). Different line search methods will be shown, including an exact line search that was the result of my research last semester. An explanation of this algorithm will be given with a description of the future work I’ll be doing this semester.
• Feb 9: Cam, Title: “Open-source tools for mapping with recreation-grade sonar: Case study on Gulf sturgeon spawning habitat in the Pearl and Pascagoula river systems” Overview: I will provide a short background on the use of recreation-grade side scan sonar (SSS) to map river substrates for the purpose of locating suitable spawning substrates for the threatened Gulf sturgeon. I will then provide an update on progress I am making on training deep learning models to automatically segment substrates from the sonar imagery including developing training datasets, hyper-parameter tuning, and preliminary results from cross validation.
• Feb 16: Karl, I will actually be doing two presentations this afternoon. The first will be a 2.5 minute scripted presentation that I am doing for $$$. It is very high level and nearly no detail will be included. Please ask questions about the mechanics and biomechanics that are unclear. This will help me out a lot. The second presentation will be the machine learning that I used on my last major project, which will form the core of my PhD defense. Title:Predicting Neuromuscular Engagement to Improve Gait Training with a Robotic Ankle Exoskeleton Overview: I will go over how I chose the model architectures to test, the study protocols, the results and the huge number of surprise challenges that changed my goals for the study. View my pre-registered hypotheses there https://osf.io/y6ghp?mode=&revisionId=&view_only=
• Feb 23: Cancelled (snow day).
• Feb 24: (special Fri meeting), 11:30AM: ROOM CHANGE: SICCS 223. Hitesh Sapotka ML prof job interview talk. Zoom: https://nau.zoom.us/j/6458547519?pwd=UGgvRU8yWkdiRmFSbzdsQjVRd0FGdz09
• Feb 27: (special Monday meeting), 10:30AM: ROOM CHANGE: SICCS 102. Rui Yu ML prof job interview talk.
• Feb 27 (special Monday meeting), 1pm: ROOM CHANGE: SICCS 102. Professor Mark Schmidt visiting from UBC CS, title: Faster Algorithms for Deep Learning? Abstract: The last 10 years have seen a revolution in stochastic gradient methods, with variance-reduced methods like SAG/SVRG provably achieving faster convergence rates than all previous methods. These methods give dramatic speedups in a variety of applications, but have had virtually no impact to the practice of training deep models. We hypothesize that this is due to the over-parameterized nature of modern deep learning models, where the models are so powerful that they could fit every training example with zero error (at least theoretically). Such over-parameterization nullifies the benefits of variance-reduced methods, because in some sense it leads to “easier” optimization problems. In this work, we present algorithms specifically designed for over-parameterized models. This leads to methods that provably achieve Nesterov acceleration, methods that automatically tune the step-size as they learn, and methods that achieve superlinear convergence with second-order information.
• Mar 2: Denae Ford Robinson, Title: The Evolution of Developers Communities, Abstract: Microsoft is home to the world’s largest developer communities and ecosystems with Azure, GitHub, and Visual Studio. Thus, having sustainable and inclusive communities is of strategic importance as it has the potential to transform society by enabling more people to develop software. Developers in these communities and others (e.g., Stack Overflow, YouTube, Twitter) often intersect aspects of their professional work with their personal life on social media platforms which allow them to feel more comfortable engaging. Therefore, understanding how developers operate at these intersections helps practitioners to better prepare for the evolution of online professional communities and continue to bridge its enterprise and consumer markets. In this talk, I will cover recent research on evolving developer communities and outline opportunities on how we can usher in the next generation of software developers by fostering healthy and inclusive communities. I will also cover the roles of researchers and research opportunities at Microsoft Research. URL for the talk: https://nau.zoom.us/my/steinmacher?pwd=R01zMWYvZWNpS1dFZitDK29pQmRtdz09 When: Thursday, March 2 at 2:20.
• Mar 9: No meeting (week before spring break).
• Mar 16: Spring Break!
• Mar 23: Daniel, Title: Cross-Validation for Testing Co-occurrence Network Inference Algorithms, Abstract: A lot of research has gone into studying microbial communities in various environments and how their interactions and relationships can provide insights into various diseases. This has led to various methods been proposed to infer microbial network interactions. Each of the methods adopt some state-of-the-art algorithm. Some of the algorithms include Correlation based techniques like the Pearson and Spearman’s Rank Correlation, regularized linear regression techniques like the Least Absolute Shrinkage and Selection Operator (LASSO), and conditional dependence techniques like the Gaussian Graphical Model. The algorithms utilize various hyper-parameters which determine the sparsity(number of edges) in network. In this research, we employ Cross-Validation Machine Learning algorithm to automatically learn such hyper-parameters rather than choosing arbitrarily. We also compare various microbial network inference algorithms in terms of their prediction accuracy when applied to some existing real microbiome data sets. Our results showed that the performance of each method varied depending on the dataset. Our research presents a comprehensive assessment and comparison of widely used microbial inference methods. The outcomes of this study can help researchers in selecting appropriate method(s) for their specific research questions, ultimately enhancing the precision and dependability of microbial data analysis.
• Mar 30: Austin - Missouri Education Predictive Modeling https://docs.google.com/presentation/d/1v3xvoQ7BdGSi2B6hNfHfckJ9G78fNKN17w56XHwwWz4/edit Sumarray of my research with the education department. Utilizing decision trees and linear models with the results and future work.
• Apr 6: Toby - Intro to deep learning in R practice talk for RezBaz AZ 2023.
• Apr 13: Danny 8 min practice talk for ASU ML Day + Jadon.
• Apr 20: Bilal - Tomorrow in the lab meeting, I will present my research objective, methodology and preliminary results. Core purpose of this presentation is to explain my approach for the candidacy exam write up, that I will be writing in summer.
• Apr 27: Karl: practice for PHD exam.
  • Paper review about over-optimistic results in the research literature / publication bias.
    • Digital medicine and the curse of dimensionality.
    • Are reported accuracies in the clinical speech machine learning literature overoptimistic?
• May 4: Daniel: proof-reading 4 page draft paper for ML in Comp Bio workshop at ICML’23.
• May 11: Toby JSM talk practice, source. Title: Efficient line search optimization of penalty functions in supervised changepoint detection. Abstract: Receiver Operating Characteristic (ROC) curves are commonly used in binary classification, and can also be used to evaluate learned penalty functions in the context of supervised changepoint detection. Since the Area Under the Curve (AUC) is a piecewise constant function of the predicted values, it can not be directly optimized by gradient descent. Recently we showed that minimizing a piecewise linear surrogate loss, AUM (Area Under Min of false positives and false negatives), results in maximizing AUC. In this talk we propose a new algorithm for AUM minimization, which exploits the piecewise linear structure to efficiently compute an exact line search, for every step of gradient descent. Because the exact line search is quadratic time in the worst case, we additionally propose an approximate line search which is log-linear time in the worst case (asymptotically the same as a constant step size). Our empirical results show that the proposed algorithm is more computationally efficient than other variants of gradient descent (constant step size, line search using grid search, etc).
• Date TBD. ROOM CHANGE: SICCS 102. LLNL postdoc Kowshik Thopalli <thopalli1@llnl.gov> invited talk Title: Improving Out-of-distribution Generalization of Deep Vision Models: Test-time and Train-time Adaptation Strategies. Abstract: This talk will focus on understanding and improving the robustness of deep vision models when training and testing data are drawn from different distributions To overcome this challenge, a common protocol has emerged, which involves adapting models at test-time with limited target data. I will first discuss a novel “align then adapt” approach that effectively exploits latent space to adapt models without requiring access to source data. This method is highly effective across multiple architectures, complex distributions shifts, and modalities. However, in cases where extremely limited target data is available, test-time approaches may fail. I will consider the extreme case where only one target reference sample is available and present our efforts in designing effective generative augmentation protocol that involves finetuning a generator with single-shot data and developing a source-free adaptation protocol with the synthetic data. Finally, in certain practical scenarios test-time adaption can be cumbersome and often we don’t have access to any target data. To address this, I will present a train-time protocol that utilizes data from multiple source domains to build generalizable computer vision models through a novel meta-learning paradigm. This approach is theoretically motivated and achieves excellent performance on more than 6 different datasets compared to a several state-of-the-art baselines. Thus, through this talk, I will present a suite of techniques to improve the out-of-distribution generalization of models for various computer vision tasks.