Data & Science with Glen Wright Colopy

Jack Fitzsimons | Evil Models: Hiding Malware in Neural Networks Did you know that it's possible to hide malware in neural networks? Actually, you can hide malware in many statistical models. This is the subject of two recently-published papers (aptly titled "EvilModel" & "EvilModel 2.0"). Dr. Jack Fitzsimons makes it easy to understand how this is done, using techniques that began long before computers. Watch or listen on... Youtube: https://youtu.be/QBnk8ogL8NkPodbean: https://dataandsciencepodcast.podbean.com/e/jack-fitzsimons-evil-models-hiding-malware-in-neural-networks/

Eric Daza | Important Ideas in Causal Inference YouTube: https://youtu.be/K5nsSMJVIT0 Andrew Gelman and Aki Vehtari wrote a paper titled, "What are the most important statistical ideas of the past 50 years?". The first idea in the list is "counterfactual causal inference". Eric Daza (Evidation Health) walks us through the main ideas of the Gelman & Vehtari paper, drawing examples from several fields, including medical & healthcare statistics. Topics0:00 - Coming up...Correlation vs Causation1:20 - Most important statistical ideas over the last 50 years6:10 - Counterfactual Causal Inference9:40 - Assumptions Change between Applied Domains21:10 - Propensity Score Methods25:15 - Transportability of Scientific Results 26:30 - People don't want generalizable results32:00 - Generic Computation Algorithms37:00 - Reweighting43:57 - Matching Methods58:20 - Medical Data is Higher Dimensional that we think.1:00:15 - Is a Trial Population Representative? 1:10:35 - Causal Models in the Future1:18:45 - Apostates Welcome1:21:45 - Scientific Debate

Ruda Zhang | Gaussian Process Subspace Regression Ruda Zhang (Duke University) walks us through "Gaussian Process Subspace Regression for Model Reduction" by Zhang, Mak, and Dunson. To keep the topic interesting for both the early career & advanced audience we recap key points at a high level so that no one gets lost. This episode involves a presentation, so you may prefer to watch the YouTube version here: https://youtu.be/IPtqUUG4XcY Ruda's website: https://ruda.city/The paper: https://arxiv.org/abs/2107.04668

Simon Mak | Integrating Science into Stats Models#statistics #science #ai It’s a common dictum that statisticians need to incorporate domain knowledge into their modeling and the interpretation of their results. But how deeply can scientific principles be embedded into statistical models? Prof. Simon Mak (Duke University) is pushing this idea to the limit by integrating fundamental physics, physiology, and biology into both the models and model inference. This includes Simon’s joint work with Profs. David Dunson and Ruda Zhang (also of Duke University). Scientific reasoning AND stats. What more could we ask for? Enjoy! Watch it on.... YouTube: https://youtu.be/bUbZO7R4z40 Podbean: https://dataandsciencepodcast.podbean.com/e/simon-mak-integrating-science-into-stats-models/ 00:00 - COMING UP….Scientists & Statisticians02:09 - Introduction - Integrating scientific knowledge into AI/ML06:08 - How much domain knowledge is sufficient?09:15 - Choosing which prior knowledge to integrate into a model14:49 - Black box & gray box optimization19:50 - Non-physics examples of integrating scientific theory into ML models22:45 - Scientific principles & modeling at different scales27:20 - Correlation is one just way of modeling linkage36:37 - Conditional independence & different-fidelity experiments39:40 - Innovation vs incorporation of known information in the model42:52 - Aortic stenosis example52:49 - Which mathematics can be used to represent scientific knowledge57:09 - How to acquire scientific domain knowledge1:02:45 - Complementary approaches to integrating science1:06:48 - Gaussian process & integrating priors over functions1:12:48 - A topic for statisticians and scientists to debate:science-based vs data-based learning. Simon Mak's Webpage: https://sites.google.com/view/simonmak/home

Dr. Jack Fitzsimons (Oblivious AI) gives a high-level introduction to the technologies that can either exploit or protect your data privacy. If you'd like to survey the landscape of data privacy-preserving technologies (from someone who's building the tech) this is a good place to start! #datascience #privacy #ai 0:00 - Coming up...3:24 - Introduction6:20 - Data privacy and privacy enhancing technologies 13:00 - History of privacy enhancing technologies19:54 - Differential privacy: Hiding the influence of a single data point22:52 - Trading data utility for data privacy38:32 - Tracking algorithms and how they decide user preferences42:04 - Preserving privacy: Anonymizing data & VPNs50:17 - Exploration vs Exploitation: Combining best of multiple domains to tackle problems54:13 - Federated learning, input and output privacy of data58:45 - Balancing data privacy vs data-driven personalization1:05:50 - What should data scientists/statisticians debate?

Chris Holmes is Professor of Biostatistics at the University of Oxford and Programme Director for Health and Medical Sciences at The Alan Turing Institute. Chris’ research interests include Bayesian nonparametrics (which is the right kind of nonparametrics), statistical machine learning, genomics, and genetic epidemiology. 0:00 - Intro1:38 - Chris Holmes, Professor of Biostatistics at Oxford University3:28 - UK Biobank & designing a valuable dataset8:42 - Healthcare charities in the UK11:16 - Digital Health: prioritizing research questions19:55 - Bayes, nonparametrics, and Bayesian nonparametrics23:30 - Model prediction is at the heart of Bayesian inference28:00 - Prioritization in model building for biology33:09 - Model constraints to generate valid inference37:34 - Hypothesis driven science in statistical learning versus deep learning43:30 - Developing models in genomics & clinical informatics48:37 - Building stable, generalizable and robust models52:41 - Important questions to think about 54:05 - Causal reasoning and clinical risk prediction57:50 - What topic should the statistical community debate?

Charlotte Deane | Bioinformatics, Deepmind's AlphaFold 2, and Llamas#datascience #ai Charlotte Deane (Oxford University) talks about statistical approaches to bioinformatics, the evolution of Google Deepmind's AlphaFold 2 & its place in protein informatics deep learning landscape. She also describes humanizing antibodies, and the increasing role of software engineers in statistical research groups. The topic of llamas, camels, and alpacas (and their unique place in proteomics research) makes a surprise visit. [Note: This episode was originally published in January 2022, but the file contained a buffering error, which prevented the full interview from being played. This version, published Feb 1, 2022 contains the full interview.] Topics0:00 Intro / An important topic to debate3:50 What is a protein? Why are proteins foundational?13:32 Immunotherapies, humanizing antibodies, & creating an scientific databases16:04 Translating in silico research into immunotherapies21:03 Nanobodies, camels, alpacas, & llamas. 25:05:00 Databases and data knowledge bases33:21:00 Targeted therapies39:45:00 Statistical modeling in proteomics45:40:00 DeepMind AlphaFold's evolution55:28:00 Software engineers in academic research groups1:03:21 The adventure of science1:07:42 Oxford Blues hockey & scientific debate

Mine Çetinkaya-Rundel | Advancing Open Access Data Science Education#datascience #statistics #education Mine Çetinkaya-Rundel (Duke University) describes the current and future states of statistics and data science education. Then she discusses the process of building open access learning material. 0:00 - Introduction1:40 - Prioritizing topics in curricula9:07 - Teaching with intent to test11:22 - Statistics without computing17:52 - What should be taught? How do we teach it?19:07 - Computational thinking is valuable (to 31:45)23:47 - Self reinforcing academics / positive feedback (to 31:45)31:08 - Data science vs statistics (the computing angle)37:55 - Statistical collaboration / technical collaboration39:45 - Common language / imputation under ignorance41:12 - Are some topics better for hands on or computational learning?45:32 - Learning computation through visualization52:40 - Video cut option before she gives an example52:42 - Let them eat cake first.56:08 - What is open source education? Open source vs open access.59:36 - Advancing open source text books1:03:55 - Economics of open source1:07:55 - The open education ecosystem1:12:17 - Modularizing & parallelizing learning topics1:16:52 - Favorite dataset on OpenIntro.Org?1:18:14 - What topic should the statistics community debate?

Gualtiero Piccinini | What Are First-Person Data? First-person methods (and its associated data) have been scientifically and philosophically contentious. Are they pseudoscientific? Or simply pushing the bounds of scientific methodology? Obviously, I have no idea… so Prof. Gualtiero Piccinini (University of Missouri – St. Louis) provides a helpful introduction to the topic covering the key points of its history and the philosophical/scientific debate. 0:00 Why cover first-person methods & data?2:26 First-person methods vs first-person data?7:10 Are first-person data legitimate at all?11:50 Phenomenology13:26 First-person data is extracted from human behavior18:25 Skepticism & arguments against first-person data25:40 Psychophysics, introspectionists, behavioralists, cognitivists, and the origins of first-person data35:20 Using new instruments & methods in science46:00 Is this where the philosophers roam? #datascience #statistics #science

Note: This conversation was recorded June 25, 2021. Martin Kuldorff | Spatiotemporal Models of OutbreaksMartin Kuldorff (Harvard Medical School) talks about the integration of biological & demographic information (and general reality) in the spatiotemporal models used to detect disease outbreaks. He also discusses how these methods can be applied to non-infectious diseases like cancer. 0:00 - Spatio-temporal modeling of outbreaks6:02 - Important features of spatio-temporal outbreak models12:20 - Which diseases wouldn't you track for modeling?19:02 - Multiple comparison adjustments of alarms25:15 - Domain knowledge of outbreak features29:30 Competing hazards & risks 34:30 Comparing hemispheres37:00 - Bridging the gap for infectious diseases to cancer45:10 - Retrospective data correction / changing monitoring 57:00 - Competing risks & statistics1:01:30 - Deducing risks & affects through knowledge of immunological mechanisms1:09:00 - Future scientific convos #datascience #science

Interesting in Data Science? Learn Data Science and Statistics from experts as they cover key topics in the field. The Data & Science podcast focusses on teaching data scientists how to think critically in order to solve data analysis problems across various scientific domains. Eric Daza | N-of-1 Science & Causal Inference | Philosophy of Data Science Much of our scientific inference revolves around the identification and replication of patterns in data. So what can be done when N=1? Eric Daza gives us a statistician's perspective on the ideas behind N-of-1 studies, its best examples, and strongest critiques. 0:00 - The purpose of N-of-1 & generalizability 3:30 - Successes and challenges in N-of-1 9:30 - A lightbulb moment 18:00 – Anomalies, Compliance, & Recurring Patterns 23:00 – Best Critiques of N-of-1, Safety, Efficacy 41:20 - Causal Inference 54:30 – Increasing the number of data scientists 1:03:30 – Biostatistics’ changing place in data science / statistical thinking

Mike Evans | Statistical Reasoning & Evidence | Philosophy of Data Science Series Mike Evans (University of Toronto) describes his approach to statistical reasoning. Mike outlines how to recognize and address problems that are statistical in nature and why these approaches should be grounded in our ability to measure statistical evidence. Watch it on YouTube at: https://youtu.be/Q7JpGZxHxXU 0:00 Statistical Reasoning2:30 The Basic Problem: Reasoning on Statistical Problems13:00 Rules of Statistical Inference19:30 Bias (The Controversial Bit?!?!)24:10 Steps of Statistical Reasoning25:50 Connection to Philosophy of Science27:35 Measuring Evidence (Frequentist vs Bayesian vs Loss Function)29:49 Problems with the p-values32:00 Choosing & Checking Priors49:25 Idealism, Good Plans, Bad Plans54:45 Describing Your Reasoning59:20 Critiques of the Principle of Evidence1:04:00 Data-Driven Science vs Hypothesis Driven Science