Leonidas Guibas, (BS ‘71, MS ‘71) Paul Pigott Professor of Computer Science at Stanford University, will discuss issues on Machine Learning.
$22 for talk and buffet lunch.
$5 for talk only. Please state “talk only” after your name when you register. You could order separately from Baylands Café.
Registration deadline: 4pm on the Monday preceding the talk. If you miss the deadline, please pay $5 at the door and you could order food from Baylands Café.
"Extracting Intelligence from Visual Data" by Leonidas Guibas (BS '71, MS '71)
For machines and autonomous agents to understand and operate in the world around us, they need to process large amounts of visual and geometric data, such as 2D images / videos, or 3D scans / models. This talk will examine the interplay of two types of computational networks that can facilitate semantic understanding and enable transfer of information and knowledge over such data.
For the first type, we have deep neural networks, which we think of as vertical networks, as they transfer information across different levels of abstraction over the same data set, from low-level features to higher-level semantic abstractions. Such networks today define the state of the art in tasks such as image classification, mapping signals (a user photo) to semantic descriptions ("contains a dog"). For the second type, which we think of as horizontal networks, they transport information between the same levels of abstraction, but over different yet related data sets, exploiting various notions of similarity. For example, articulation functionality can be transferred from a pair of scissors to a pair of pliers, because of their structural resemblance. Such horizontal networks can be built using functional maps, which are linear operators transferring knowledge-encoding functions between connected data sets.
I will briefly discuss some of the issues involved on the construction of both types of networks, especially regarding 3D data that are normally represented in irregular formats such as point clouds or meshes, as opposed to the regular image grids that facilitate the common convolutional neural networks in the 2D image world.
Both types of networks encode data in certain latent representations that are not well understood. Though normally obtained independently, these encodings have in fact many important inter-relations. In the end we want both vertical and horizontal networks to "play well" with each other, giving rise to commutative map diagrams that enforce structure-preserving abstractions, making these encodings consistent with each other. The result is a greatly reduced supervision burden on these networks without the danger of over-fitting.
I will demonstrate these machine learning ideas in the context of image and shape classification and segmentation, 3D reconstruction from 2D data (e.g., creating a 3D model of an object from just one image of the object), and other visual tasks.
About our Speaker
Leo (Caltech B.S. & M.S.) obtained his Ph.D. from Stanford University under the supervision of Donald Knuth. His main subsequent employers were Xerox PARC, DEC/SRC, MIT, and Stanford. He is currently the Paul Pigott Professor of Computer Science (and by courtesy, Electrical Engineering) at Stanford University and heads the Geometric Computation group in the Computer Science Department of Stanford University.
Leo is a member and past acting director of the Artificial Intelligence Laboratory and a member of the Computer Graphics Laboratory, the Institute for Computational and Mathematical Engineering (iCME) and the Bio-X program. He has been elected to the US National Academy of Engineering and the American Academy of Arts and Sciences, and is an ACM Fellow, an IEEE Fellow and winner of the ACM Allen Newell award and the ICCV Helmholtz prize.