BEGIN:VCALENDAR
VERSION:2.0
PRODID:-//UM//UM*Events//EN
CALSCALE:GREGORIAN
BEGIN:VTIMEZONE
TZID:America/Detroit
TZURL:http://tzurl.org/zoneinfo/America/Detroit
X-LIC-LOCATION:America/Detroit
BEGIN:DAYLIGHT
TZOFFSETFROM:-0500
TZOFFSETTO:-0400
TZNAME:EDT
DTSTART:20070311T020000
RRULE:FREQ=YEARLY;BYMONTH=3;BYDAY=2SU
END:DAYLIGHT
BEGIN:STANDARD
TZOFFSETFROM:-0400
TZOFFSETTO:-0500
TZNAME:EST
DTSTART:20071104T020000
RRULE:FREQ=YEARLY;BYMONTH=11;BYDAY=1SU
END:STANDARD
END:VTIMEZONE
BEGIN:VEVENT
DTSTAMP:20260828T094139
DTSTART;TZID=America/Detroit:20260911T100000
DTEND;TZID=America/Detroit:20260911T110000
SUMMARY:Workshop / Seminar:When Prediction Meets Randomization: Regression\, Machine Learning\, or LLMs? Lessons from 125 Randomized Trials
DESCRIPTION:Covariate adjustment is a standard tool for improving precision in randomized trials\, but the value of increasingly flexible prediction methods remains unclear. This talk begins with an empirical comparison of regression and machine-learning-based adjustment across 50 completed trials. The results show that flexible machine learning methods are not automatically superior: simple regression adjustment with prognostic baseline covariates is often highly competitive. This finding motivates a new question: can large language models change the picture by extracting useful prognostic information from baseline covariates\, trial descriptions\, or other text-derived features? To address this question\, I present a unified framework for LLM-assisted covariate adjustment and empirical evidence from 125 randomized trials. We evaluate LLM-derived features\, including zero/few-shot predictions\, fine-tuned models\, and embedding-based representations\, combined with both regression and machine-learning estimators. The results show that LLM features can provide additional precision gains beyond classical covariates\, although the gains vary across trials\, sample sizes\, and outcome types. Overall\, LLMs appear most useful as tools for constructing better prognostic features within transparent and honest adjustment procedures.\n\n*Department reception in 450 West Hall after the seminar at 11:00 AM*
UID:151079-21910841@events.umich.edu
URL:https://events.umich.edu/event/151079
CLASS:PUBLIC
STATUS:CONFIRMED
CATEGORIES:seminar,statistics
LOCATION:West Hall - 340
CONTACT:
END:VEVENT
BEGIN:VEVENT
DTSTAMP:20260904T121124
DTSTART;TZID=America/Detroit:20260918T100000
DTEND;TZID=America/Detroit:20260918T110000
SUMMARY:Workshop / Seminar:MonarchAttention: Zero-Shot Conversion to Fast Structured Attention
DESCRIPTION:Transformers have achieved state-of-the-art performance across various tasks\, but suffer from a notable quadratic complexity in sequence length due to the attention mechanism. In this talk\, I will share our work on MonarchAttention – a novel approach to sub-quadratic attention approximation via Monarch matrices\, an expressive class of structured matrices. \nBased on the variational form of softmax\, we describe an efficient optimization-based algorithm to compute an approximate projection of softmax attention onto the class of Monarch matrices with Θ(N √N d) computational complexity and Θ(N d) memory and IO complexity. We show with experiments that unlike previous approaches\, MonarchAttention is both (1) transferable\, yielding minimal performance loss with no additional training\, even when replacing every attention layer of the transformer\, and (2) hardware-efficient\, utilizing the highest-throughput tensor core units on modern GPUs. With optimized kernels\, MonarchAttention achieves substantial speed-ups in wall clock time over FlashAttention-2: 1.4× for shorter sequences (N = 256) and 8.2× for longer sequences (N = 16K). \nWe demonstrate the quality of MonarchAttention on diverse tasks and architectures in vision and language problems\, showing that it flexibly and accurately approximates softmax attention in a variety of contexts. This work is with Can Yaras\, Alec Xu\, Pierre Abillama\, and Changwoo Lee. Our code is available at https://github.com/cjyaras/monarch-attention.
UID:151572-21911779@events.umich.edu
URL:https://events.umich.edu/event/151572
CLASS:PUBLIC
STATUS:CONFIRMED
CATEGORIES:seminar,Statistics,Talk
LOCATION:West Hall - 340
CONTACT:
END:VEVENT
END:VCALENDAR