Inductive Bias in Deep Learning: From Structure to Training

Share

Machine learning requires assumptions and data, with the assumptions based on domain knowledge. With the exponential increase in dataset set and computation we’ve been able to reduce the domain expertise and inductive bias we integrate in computer vision models, from manually designed features such as SIFT, to learning representations with deep learning. Recently this approach has been pushed even further, with Convolutional Neural Networks (CNNs), integrating strong inductive bias in their structure, being replaced with Vision Transformers (ViT), or even fully-connected neural networks with alternative training regimes. In this talk, Dr. Yanni Ioannou, Assistant professor at the University of Calgary, will discuss the current limits of this approach, and the need for more domain-agnostic learning.

 


  Date and Time

  Location

  Hosts

  Registration



  • Date: 20 Apr 2022
  • Time: 03:00 PM to 04:00 PM
  • All times are (GMT-07:00) America/Edmonton
  • Add_To_Calendar_icon Add Event to Calendar

A Zoom invitation link will be sent to registrants prior to event time.

  • Calgary, Alberta
  • Canada
  • Building: Virtual

  • Starts 21 March 2022 10:00 AM
  • Ends 20 April 2022 12:00 PM
  • All times are (GMT-07:00) America/Edmonton
  • No Admission Charge


  Speakers

Yani Ioannou Yani Ioannou

Email:

Address:ICT248, Information and Communications Technology Building, University of Calgary, , Calgary, Alberta, Canada