Place: Online
Date: 9-10 September 2020
Time: 9:00 - 12:00 (CEST)
Description
In caption generator systems, the identification of the figures depicted in an image depends on a matrix of weights for each of the classes with which the system has been trained. Current caption generators are trained with images that reflect present time lifestyles. Therefore the matrix is in fact a TimeMatrix of the present. The course will show how the identification of classes varies as the matrix of weights depends on data relative to past centuries. This produces a time machine effect where the bike of a person in the TimeMatrix of the present becomes a horse of Saint George in the TimeMatrix of the 15th century.
This webinar will demonstrate the SGoaB results and potential to adapt automatically produced descriptions of paintings to the time period when they were created. The course will pose and discuss challenges for researchers. At the end of the course, a demo will be conducted to show correction of anachronisms and class refinement examples.
Learning outcomes
Participants will be introduced to what we call the time machine effect, which consists of the objects of an image being transformed via deep learning methods to similar concepts that are more appropriate to another time period. The technical challenges and current solutions will be discussed.
Target group
Academic researchers in data mining and data science that are interested in art and culture.
Prerequisites
Active interest in cultural heritage a must.
Basic knowledge of Natural Language Processing a plus.
Using Neural Networks a plus.
Course lecturers
Maria Cristina Marinescu (CASE Department, Barcelona Supercomputing Center)
Joaquim More Lopez (CASE Department, Barcelona Supercomputing Center)
Artem Reshetnikov (CASE Department, Barcelona Supercomputing Center)
Agenda
Day 1. Understanding the past through its parallels with the present
| 9:00-9:20 | Introduction to the general issue of describing cultural heritage images: what is done, and what is used for? | Maria Cristina Marinescu |
| 9:20-10:05 | AI-powered semantic labelling of image datasets- formats, target concepts, challenges and description of TimeMatrix and its purpose | Joaquim More Lopez |
| 10:05-10:35 | Training data in image datasets - images and metadata, challenges | Artem Reshetnikov |
| 10:35-10:45 | Break | |
| 10:45-11:10 |
Present state-of-the-art caption generators and their limitations when dealing with anachronic images |
Joaquim More Lopez |
| 11:10-11:45 | Transfer learning to describe cultural heritage images as they differ from the present way of life - anachronisms, evolution of concepts through time, minorities, symbols | Artem Reshetnikov |
| 11:45-12:00 |
Wrap-up + Connecting the dots to see the bigger picture |
Maria Cristina Marinescu |
Day 2. Technical implementation of the TimeMatrix
| 9:00-9:15 | Filtering anachronic classes | Joaquim More Lopez |
| 9:15-9:45 | Defining new visual classes and visual relations for cultural and iconographic symbols | Joaquim More Lopez |
| 9:45-10:15 | Refining the identification of classes and visual relations according to time contex | Artem Reshetnikov |
| 10:15-10:30 | Deep Learning implementation of anachronism correction and class refinement | Artem Reshetnikov |
| 10:30-10:40 | Break | |
| 10:40-11:00 | Current state of the Saint George on a Bike project | Joaquim More Lopez |
| 11:00-11:30 | Class features and Deep Learning implementation | Artem Reshetnikov |
| 11:30-12:00 | Discussion of anachronic correction and class refinement examples (demo) | Artem Reshetnikov |
Registration is closed.
