Talk on For Lack of a Better Word: Mining Visual Variation Behind Labels

Title 

For Lack of a Better Word: Mining Visual Variation Behind Labels

Abstract

Learning to assign the continuum of our visual world into discrete categories is a fundamental way knowledge is produced and reproduced.
Yet even in established categories, like the names of animals, distinct variation may loom large. Broadly, it occupies two different scales:

1. On a small scale, it concerns minute details that are often impossible to reliably describe or communicate through language.
   One such case is the morphological evolution of hand-written characters studied by the field of palaeography.
2. On a large scale, it clusters into visual vocabularies that lie behind ordinary concepts such as the name of a country, a decade, or a scene.
   Almost ten years ago, research interest in mining and organizing such variation culminated, by exhausting the technics of its time [1, 2].
   Ten years later, we try to spark it back by building on recent synthesis methods that learn elaborate compositional representations of their training data.

Presented work:
Part I. learnable-typewriter.github.io, learnable-typewriter-pal.github.io
Part II. diff-mining.github.io

References:
[1] aclanthology.org/P13-1021.pdf
[2] graphics.cs.cmu.edu/projects/whatMakesParis

 

Bio

Yannis Siglidis (he/they), is a final year PhD-student in Computer Vision advised by Mathieu Aubry at the Imagine Lab in Paris.
Formally trained in Computer Science and Machine Learning, Yannis has also performed interdisciplinary research in digital sociology, AI art and philosophy.
More info: ysig.github.io

 

Sign-up is required to attend this talk.