BEGIN:VCALENDAR
VERSION:2.0
PRODID:-//Talks.cam//talks.cam.ac.uk//
X-WR-CALNAME:Talks.cam
BEGIN:VEVENT
SUMMARY:Analysis and Applications of Deep Cascade Learning - Doris Xin Du
DTSTART:20220610T120000Z
DTEND:20220610T130000Z
UID:TALK175544@talks.cam.ac.uk
CONTACT:Yuan Huang
DESCRIPTION:This talk is on the analysis and applications of a constructiv
 e architecture for training Deep Neural Networks (DNNs)\, which are usuall
 y trained by End-to-End (E2E) gradient propagation with fixed depths. E2E 
 training of DNNs has proven to offer impressive performances in a number o
 f applications such as computer vision\, machine translation and in playin
 g complex games such as GO. However\, the massive cost in computing and me
 mory hinders its applications in many areas\, such as portable medical dev
 ices. Moreover\, the majority of DNNs are data hungry which raises further
  barriers of applications in regions where data collection or labelling is
  expensive. As an alternative\, Cascade Learning (CL)\, the approach of in
 terest here\, trains networks in a layer-wise fashion and has been demonst
 rated to achieve satisfactory performance in large scale tasks such as the
  popular ImageNet benchmark dataset\, at substantially reduced computing a
 nd memory requirements. Here we focus on the nature of features extracted 
 from CL. By attempting to explain the process of learning using the Inform
 ation Bottleneck theory\, an empirical rule (Information Transition Ratio)
  is derived to automatically determine a satisfactory depth for Deep Neura
 l Networks. We suggest that CL packs information in a hierarchical manner\
 , with coarse features in early layers and more task specific features in 
 later layers. This is verified by considering Transfer Learning whereby fe
 atures learned from a data-rich source domain assist in learning a data-sp
 arse target domain. Using a wide range of inference problems in medical im
 aging\, human activity recognition and inference from single cell gene exp
 ression between mice and humans\, Transfer Learning from a cascade trained
  model shows significant advantages in small data regime. \n\nThe seminar 
 will be held in a hybrid format. We strongly encourage you to participate 
 in person at MR 2\, Centre of Mathematical Sciences\, CB3 0WA. Althernativ
 ely you can join using the following Zoom link:\n\n*Join Zoom Meeting*\nht
 tps://maths-cam-ac-uk.zoom.us/j/99801605037?pwd=bnlTanhLUVRKenBmTGNYckR4dn
 I1QT09\n\nMeeting ID: 998 0160 5037\n\nSecurity Passcode:  435654
LOCATION:This talk will be held in a hybrid format at MR 2\, Centre of Mat
 hematical Sciences\, CB3 0WA. Alternatively you can also join with Zoom (s
 ee abstract for Zoom link).
END:VEVENT
END:VCALENDAR