- This event has passed.
ECSS Symposium – August 2019
September 17, 2019 @ 1:00 pm - 2:00 pm EDT
Please join us for our next ECSS Symposium.
Tuesday, September 17, 10:00 AM Pacific / 1:00 PM Eastern
Join from PC, Mac, Linux, iOS or Android: https://zoom.us/j/114343187
Or iPhone one-tap : US: +16468769923,,114343187## or +16699006833,,114343187##
Or Telephone: Dial(for higher quality, dial a number based on your current location): US: +1 646 876 9923 or +1 669 900 6833 or +1 408 638 0968
Meeting ID: 114 343 187
International numbers available: https://zoom.us/zoomconference?m=Kzg5xp93d84wOReXpCZzGCWHpKACTSLh
More information about the Symposium series and links to previous presentation can be found here
Robert Sinkovits, Ph.D.
Co-director XSEDE ECSS
San Diego Supercomputer Center
Paul Rodriguez (SDSC)
The “Morelli Machine”: A Proposal Testing a Critical, Algorithmic Approach to Art History
The Morelli Machine refers to an algorithmic approach to characterizing authorship from the late 19th century which proposed that fine details of minor items in a painting would reveal particular styles. The PIs set out to test the hypothesis that contemporary computer vision techniques could perform this sort of “stylistic” matching. In order to do this, they sought to mechanize a method that is indigenous to art history and that uses details as a proxy for style. This project approached the question of “style” as one of extracting features that have some discriminatory power for distinguishing paintings or groups of paintings. We used feature discovery from a pretrained convolution network (VGG19) for object recognition. We processed both whole images and some class of image parts (ie mouths), and performed clustering. In this presentation I will review the image preparation steps, extraction steps, clustering results, and cluster evaluation. The upshot is that all convolution layers indeed have discriminatory features, and different layers might have different kinds of features, with different interpretability that may be hard to define.