BEGIN:VCALENDAR
VERSION:2.0
PRODID:-//Talks.cam//talks.cam.ac.uk//
X-WR-CALNAME:Talks.cam
BEGIN:VEVENT
SUMMARY:Long-tailed Recognition via Key Attribute Learning - Yu Fu (Aberys
 twyth University)
DTSTART:20240701T150000Z
DTEND:20240701T151000Z
UID:TALK217483@talks.cam.ac.uk
DESCRIPTION:Deep learning models often struggle with datasets exhibiting l
 ong-tailed distributions\, where the majority of data is concentrated in a
  few categories\, leaving many with very few samples. This imbalance resul
 ts in models favouring well-represented categories\, leading to poorer per
 formance for those with fewer instances. Existing methodologies focus on a
 ddressing class-wise imbalance but disregard the attribute-wise disparitie
 s. By assigning equal weight to each instance within a class\, these appro
 aches overlook the long-tailed distribution of attributes\, thus underrepr
 esenting information from infrequent attributes. The reduction in feature 
 diversity consequently diminishes model performance. To address this chall
 enge\, we introduce an innovative methodology\, namely Key Attribute Learn
 ing (KAL). It emphasises the importance of less common attributes by utili
 sing the Instance Diversity Index (IDI) to assess and prioritise attribute
  diversity for each instance. KAL effectively expands feature margins amon
 g categories and addresses the overfitting problem. Our results demonstrat
 e that KAL is non-invasive in both single-model and Mixture of Experts (Mo
 E) settings. Implementing our method on BalPoE\, we attained state-of-the-
 art (SOTA) performance on CIFAR-100-im100\, ImageNet-LT\, and iNaturalist 
 datasets\, showcasing its broad applicability and significant improvements
  across both balanced and diverse test distributions.\nCo-Author: Jungong 
 Han*\, Xiang Chang\, Changrui Chen\, Changjing Shang\, Qiang Shen\n&nbsp\;
LOCATION:Seminar Room 1\, Newton Institute
END:VEVENT
END:VCALENDAR