The workshop “Learning Data Representation: Hierarchies and Invariance” was held at the McGovern Institute for Brain Research at MIT (Bldg. 46), from November 22-24, 2013.
The goal of the meeting is to investigate advances and challenges in learning “good representations” from data, in particular representations that can reduce the complexity of later supervised learning stages. The meeting will gather experts in the field to discuss current and future challenges for the theory and applications of learning representations.
This workshop was organized by Tomaso Poggio and Lorenzo Rosasco and sponsored by the Laboratory for Computational and Statistical Learning (LCSL), joint between the Istituto Italiano di Tecnologia and Massachusetts Institute of Technology and the Center for Brains, Minds and Machines (CBMM).
Speakers have kindly permitted us to share their presentation slides along with videos of most of the talks. Please note a few talks are unavailable due to technical difficulties experienced with the AV equipment during the event.
Friday November 22, 2013
Introduction to the workshop
Tomaso Poggio, Director of the Center for Brains, Minds and Machines (CBMM), presented an introduction to CBMM and introduced workshop organizer Dr. Lorenzo Rosasco.
Lorenzo Rosasco (LCSL, MIT) welcomes guests to the Workshop on Learning Data Representation: Hierarchies and Invariance.
Session 1: Early Features in Vision
Chair: Carlo Tomasi (Duke U.)
Panel discussion: Early Features in Vision
Moderator: Carlo Tomasi (Duke U.)
Panel: William Freeman (MIT CSAIL), Charles Stevens (Salk Institute), Thomas Serre (Brown University), Giulio Sandini (IIT), and Pietro Perona (Caltech.)
Saturday November 23, 2013
Session 2: Learning Features and Representations
Moderator: Alessandro Verri (Università degli Studi di Genova)
Panel: Benjamin Recht (UC Berkeley), Lorenzo Rosasco (LCSL, MIT), Misha Belkin (Ohio State), Joachim Buhmann (ETH Zurich), and Guillermo Sapiro (Duke).
Session 3: Learning Invariances and Hierarchies
Chair: James DiCarlo (MIT)
Scattering bricks to build invariants for perception (part 1), Stéphane Mallat (École Polytechnique)
Moderator: James DiCarlo (MIT)
Panel: Yann LeCun (NYU), Stéphane Mallat (École Polytechnique), Tomaso Poggio (MIT), Pierre Baldi (U.C.Irvine), and Maximilian Riesenhuber (Georgetown).
Note: Maximilian Riesenhuber (Georgetown) participated in this panel. Unfortunately, the video and slides of his talk are not available.
Sunday November 24, 2013
Session 4: Beyond Feedforward Architectures
Chair: Josh Tenenbaum (MIT, CBMM)
Compositional models: complexity of representation and inference, Alan Yuille (UCLA)
Panel: Stefano Soatto (UCLA), Shimon Ulllman (MIT/Weizmann Inst.), Alan Yuille (UCLA), and Russ Salakhutdinoff (U. Toronto).
What Makes a Good Representation? From Invariance to Causality, Josh Tenenbaum (MIT, CBMM)