BEGIN:VCALENDAR
VERSION:2.0
PRODID:-//Date iCal//NONSGML kigkonsult.se iCalcreator 2.20.2//
METHOD:PUBLISH
X-WR-CALNAME;VALUE=TEXT:Eventi DIAG
BEGIN:VTIMEZONE
TZID:Europe/Paris
BEGIN:STANDARD
DTSTART:20231029T030000
TZOFFSETFROM:+0200
TZOFFSETTO:+0100
RDATE:20241027T030000
TZNAME:CET
END:STANDARD
BEGIN:DAYLIGHT
DTSTART:20240331T020000
TZOFFSETFROM:+0100
TZOFFSETTO:+0200
TZNAME:CEST
END:DAYLIGHT
END:VTIMEZONE
BEGIN:VEVENT
UID:calendar.28067.field_data.0@www.u-gov-ricerca.uniroma1.it
DTSTAMP:20260411T201319Z
CREATED:20240506T090534Z
DESCRIPTION:Abstract: Deep learning\, and in particular\, large language mo
 dels have made great strides in many fields including vision\, language\, 
 and medicine. The impressive performance of large models comes at a signif
 icant price: the models tend to be billions to trillions of parameters in 
 size\, are expensive to train\, have a huge operational cost\, and typical
 ly need cloud service for deployment. Meanwhile\, considerable research ef
 forts have been devoted  to designing smaller/cheaper models\, at the pric
 e of restricted generalizability and performance. Not all queries we may w
 ish to pose to a model are hard. Some queries can be answered nearly as ac
 curately with cheaper models at a fraction of the cost of the larger model
 s. However\, the performance of cheaper models may suffer on other queries
 . Can we combine the best of both worlds by striking a balance between cos
 t and performance? In this talk\, I will describe two settings in which ou
 r group has tackled this issue.In the first setting\, we are interested in
  approximate answers to queries over model predictions. We show how\, unde
 r some assumptions about the cheap model\, queries can be answered with a 
 provably high precision or recall by using a judicious combination of invo
 king the large model on data samples and the cheap model on data objects. 
 In the second setting\, we are interested in learning a router\, which\, g
 iven a query\, predicts its level of hardness\, based on which the query i
 s either routed to the small model or to the large model. For both setting
 s\, results of extensive experiments show the effectiveness and efficiency
  of our approach. Speaker's Bio: Laks V.S. Lakshmanan is a professor of Co
 mputer science at UBC\, Vancouver\, Canada. His research interests span a 
 wide spectrum of topics in data management\, integration\, cleaning\, and 
 warehousing\; data mining\; semi-structured and unstructured data\; big gr
 aphs\, social networks and social media\; ML\, NLP\; and efficient deep le
 arning. He is an ACM Distinguished Scientist and has won several awards in
 cluding best paper awards and distinguished reviewer awards. He has served
  on most top conferences and journals in his areas of research\, on progra
 m committees\, as senior PC member\, meta-reviewer\, general chair\, and a
 s associate editor. 
DTSTART;TZID=Europe/Paris:20240516T143000
DTEND;TZID=Europe/Paris:20240516T143000
LAST-MODIFIED:20240506T101610Z
LOCATION:Aula Magna DIAG
SUMMARY:Taming the Cost of Deep Neural Models: Hybrid Models to the Rescue?
  - Laks V.S. Lakshmanan
URL;TYPE=URI:http://www.u-gov-ricerca.uniroma1.it/node/28067
END:VEVENT
END:VCALENDAR