@max_spero_ I'm curious what you think the value add of Pangram is if all LLM companies embed watermarks in their text and image outputs. Is it to detect mixed authorship, or the degree to which the authorship is mixed?
Spotlight: We propose a survival analysis method for settings where censoring is driven by interventions whose prevalence shifts over time. By @MeeraKrishnamo1, Donna Tjandra, & Jenna Wiens from @Michigan_AI; Amanda Kowalski from @umichECON; and @dmshanmugam from @cornell_tech.
๐ฃ๏ธ My oral presentation: Tues 1โ3 PM in Room 801B
๐ My poster presentation: Tues 5:30โ8:30 PM
๐ Link to our paper: https://t.co/fyb4MuW0lX
Excited to chat with others working on robust ML!
Excited to be at #KDD25 this week sharing our work on robust ML! This work was inspired by a real challenge my collaborators, Michael Sjoding and Jenna Wiens, faced when building a model for Michigan Medicine during the height of the COVID-19 pandemic. ๐งต
We propose a new CV-based method that more reliably selects models that rely on stable, as opposed to unstable correlationsโleading to more reliable model performance over time.
How can #AI help stop hospital-acquired infections?
This cross-disciplinary effort, including Prof. Jenna Wiens, brings together computer science, clinical care, & data science to reduce C. diff spread w/ a real-world deployed AI tool. @UMengineering
๐ฐ https://t.co/hKvZTZcNXS
(1/) Excited to announce our new publication on biases in lab testing and its implications for AI in healthcare! Read more from @UMengineering: https://t.co/KT65wb8vgu
Many MIL tasks contain absolute position information, e.g., pathological findings in chest x-rays. Standard (non-transformer) MIL approaches are fast, but do not leverage this position information. Transformer-based MIL approaches leverage position information, but are slow.
This is commonly used to classify large images. To avoid information loss that comes with down sampling these large images, one can instead divide large images into patches that maintain resolution, and then classify bag of patches
We consider the multiple instance learning (MIL) setting. As opposed to standard supervised learning (mapping one input to one output), in MIL, a bag of inputs is mapped to a single output.
In #healthcare, many #AI models can be used in settings other than those for which they were approved.
In this recent Nature Medicine article, @MeeraKrishnamo1, M. Sjoding & Jenna Wiens focus on such off-label use of #AI and ways to ensure patient safety:
https://t.co/LhebyAPYMT
1/ Can AI explanations help clinicians catch when AI is biased? We answer this in our new work published in JAMA @JAMA_current!
https://t.co/ataSuZ5Mkx