Archive of ‘Seminar’ category
When/where: 12pm/E2-215 (pizza will be provided!)
Presenter: Erica Rutter (UC Merced; Applied Math)
Title: Methods for Few-Shot Biomedical Image Segmentation and Learning Equations from Data
Abstract: Machine learning methods have had a powerful impact in the field of biomedical image segmentation and analysis. Traditionally, these machine learning methods have had two drawbacks: they do not ensure contiguity of a segmented object, and they require many expensive manually annotated images to train. I will present a novel methodology for image segmentation by reformulating the machine learning method to trace the boundary of an object, the way a human would. We explore the accuracy of the method on benchmark cell segmentation datasets as well as in-house ‘real’ data. We further compare the ability of the method to work in the low-data limit and show that our method can be used to generate training data for more sophisticated segmentation methods, thereby reducing the burden of human annotation. I will then explore a hybrid framework the combines elements from dynamical systems and machine learning to analyze the quantifiable data generated from such image segmentations. In particular, I will present a robust method for learning underlying dynamical systems from noisy spatiotemporal data, drawing from examples in cell migration and cancer.
When/where: 12pm/E2-215 (pizza will be provided!)
Presenter: Vanessa Boehm (LBNL)
Title: The powers and pitfalls of deep generative models in scientific applications
Abstract: Of all machine learning methods generative models are particularly interesting for scientific applications because of their probabilistic nature and ability to fit complex data and probability distributions. However, in their vanilla form generative models have a number of shortcomings and failure modes which can be a hindrance to their application: They can be difficult to train on high dimensional data and they can fail in crucial tasks such as outlier detection, correct uncertainty estimation or the generation of realistic artificial data. In my talk I am going to explore the reasons for these failures and propose new generative models and generative model based approaches that are robust to these shortcomings. The proposed approaches are easy to train and validate, numerically stable and do not require fine-tuning. They should thus be particularly fitting for scientific applications. I will demonstrate how these approaches can be used for scientifically relevant tasks such as realistic data generation, probabilistic inference on corrupted data and outlier detection.
Where/when: E2-215 (pizza will be provided!)
Presenter: David Shih (Rutgers University)
Title: “New Approaches to Anomaly Detection Inspired by High Energy Physics”
Abstract: With an enormous and complex dataset, and the ability to cheaply generate realistic simulations, the Large Hadron Collider (LHC) provides a novel arena for machine learning and artificial intelligence. In this talk I will give an overview of a number of recently proposed methods for anomaly detection that were motivated by the search for new particles and forces at the LHC. This includes methods based on autoencoders, weak supervision, density estimation and simulation-assisted reweighting. I will also summarize the status of the ongoing “LHC Olympics 2020” anomaly detection data challenge, where many of these techniques are being applied to “black box” datasets by a number of groups from around the world. Finally, while these methods are being developed in the context of high energy physics, I will attempt to highlight the ways in which they are general and could be applied to data-driven discovery in other branches of science and beyond.
When/Where: E2-215 at 12pm on January 22nd, 2020
Presenter: Ioannis “Yianni” Anastopoulos (UC Santa Cruz)
Title: Generalizing Drug Response from Cell Lines to Patients
Abstract:
Cancer treatment poses a unique challenge in part due to the heterogeneous nature of the disease. It is estimated that a staggering 75% of cancer patients do not respond to any type of cancer treatment. Preclinical models, such as cell lines and patient-derived xenografts, have been used to better understand the genes and pathways that contribute to tumorigenesis, as well as to identify markers of treatment response. For my thesis project, I am leveraging deep learning techniques to generalize drug response prediction from preclinical models to patients.
Standard machine learning methods have modest prediction accuracy and are challenged by the ever-increasing dimensionality of available data. Deep learning solves many shortcomings of previous methods and has been used successfully to advance drug design, drug-target profiling, and drug repositioning. In my preliminary results, I have shown that a deep learning model incorporating transcriptome and drug structure information achieves competitive drug response prediction performance on the Cancer Cell Line Encyclopedia (CCLE) dataset. I plan to extend this work to enable generalization to future drug compounds.
We kick off the New Year with a seminar by at 12pm in E2-215 or
by Zoom given by:
Speaker: Rich Caruana (Microsoft)
on
Title: Friends Don’t Let Friends Use Black-Box Models: The Importance of Interpretability in Machine Learning
Abstract: Every data set is flawed, often in ways that are unanticipated and difficult to detect. If you can’t understand what your model learned from the data, your model probably is less accurate than it could be, and might even be risky to deploy. Unfortunately, historically there has been a tradeoff between accuracy and intelligibility: accurate models such as deep neural nets, boosted tress and random forests are not very intelligible, and intelligible models such as linear regression and decision trees usually are less accurate. In mission-critical domains such as healthcare, where being able to understand, validate, edit and ultimately trust a model is important, one often had to choose less accurate models. But this is changing. We have developed a learning method based on generalized additive models with pairwise interactions (GA2Ms) that is as accurate as full complexity models, yet even more interpretable than linear regression. In this talk I’ll show how interpretable, high-accuracy machine learning is helping us discover what our models learned and uncover flaws lurking in our data. I’ll also show how we’re using these models to uncover bias where fairness and transparency are important. Code for GA2Ms is available at https://github.com/interpretML.
Bio: Rich Caruana is a Senior Principal Researcher at Microsoft. His research focus is on intelligible/transparent modeling, machine learning for medical decision making and computational ecology, and deep learning. Before joining Microsoft, Rich was on the faculty in Computer Science at Cornell, at UCLA’s Medical School, and at CMU’s Center for Learning and Discovery. Rich’s Ph.D. is from CMU, where he worked with Tom Mitchell and Herb Simon. His thesis on Multitask Learning helped create interest in the subfield of machine learning called Transfer Learning. Rich received an NSF CAREER Award in 2004 (for Meta Clustering), best paper awards in 2005 (with Alex Niculescu-Mizil), 2007 (with Daria Sorokina), and 2014 (with Todd Kulesza, Saleema Amershi, Danyel Fisher, and Denis Charles), and co-chaired KDD in 2007 with Xindong Wu.
Title: Hidden Physics Models: Machine Learning of Non-Linear Partial Differential Equations
Speaker: Maziar Raissi (Brown University)
Abstract: A grand challenge with great opportunities is to develop a coherent framework that enables blending conservation laws, physical principles, and/or phenomenological behaviors expressed by differential equations with the vast data sets available in many fields of engineering, science, and technology. At the intersection of probabilistic machine learning, deep learning, and scientific computations, this work is pursuing the overall vision to establish promising new directions for harnessing the long-standing developments of classical methods in applied mathematics and mathematical physics to design learning machines with the ability to operate in complex domains without requiring large quantities of data. To materialize this vision, this work is exploring two complementary directions: (1) designing data-efficient learning machines capable of leveraging the underlying laws of physics, expressed by time dependent and non-linear differential equations, to extract patterns from high-dimensional data generated from experiments, and (2) designing novel numerical algorithms that can seamlessly blend equations and noisy multi-fidelity data, infer latent quantities of interest (e.g., the solution to a differential equation), and naturally quantify uncertainty in computations. The latter is aligned in spirit with the emerging field of probabilistic numerics.
@UCSC: E2-215 at 12pm PT
Online: https://ucsc.zoom.us/j/468337241
Title: TRIMAP: LARGE-SCALE DIMENSIONALITY REDUCTION USING TRIPLETS
Speakers: Ehsan Amid and Manfred Warmuth (UCSC / visiting Google Brain)
Abstract: We introduce “TriMap”; a dimensionality reduction technique based on triplet constraints that preserves the global accuracy of the data better than the other commonly used methods such as t-SNE, LargeVis, and UMAP. To quantify the global accuracy, we introduce a score which roughly reflects the relative placement of the clusters rather than the individual points. We empirically show the excellent performance of TriMap on a large variety of datasets in terms of the quality of the embedding as well as the runtime. On our performance benchmarks, TriMap easily scales to millions of points without depleting the memory and clearly outperforms t-SNE, LargeVis, and UMAP in terms of runtime.
@UCSC: E2-215 at 12pm PT
Online: https://ucsc.zoom.us/j/120429746
When/where: E2-215 at 12pm
Presenter: Oskar Elek
Title: Monte Carlo Physarum Machine: Unconventional AI for astronomy and beyond
Abstract: Slime mold (Physarum Polycephallum) is a freak of the natural world that – out of decaying forest debris – builds near-optimal transport networks. We leverage a custom Monte-Carlo simulation of this organism to approximate such transport networks: both in terms of structure and their density (likelihood) in 3D space. I will discuss our analysis of the Cosmic Web using this hybrid model, as well as future prospects in training and adapting it to several other domains.
With luck this Zoom link will work.
When/where: E2-215 at 12pm
Presenter: Reuben Harry Cohn-Gordon (Stanford University)
Title: Bayesian Models of Pragmatics for Natural Language
Abstract: Emerging from work in Bayesian cognitive science and game theory, probabilistic models of pragmatic reasoning have been successful at modelling human inferences about linguistic meaning, on the basis of their interlocutor’s intentions and knowledge. However, they have largely been applied to idealized domains. Meanwhile, natural language processing systems have made significant progress at open-domain tasks requiring language understanding, but often struggle to behave in human-like ways. I describe work combining these approaches, with the aim of obtaining interpretable models of pragmatic reasoning for natural language.
Where: CfAO (we return to E2-215 next week)
When: 12-1pm
What: Or own David Reiman (Physics) will describe his internship over the Summer at Deep Mind. Others are encouraged to the fill the remainder of the time by describing their own AI-related activities over the summer.
We will have a formal seminar on October 23, 2020 when we return to E2-215. And it rumored that pizza will appear.