Title: Deep Convolution Networks as Geometric Image Representations

Until recently, computer vision algorithms mostly resulted from geometric considerations over shapes, textures, motion and 3D perspective projections. Deep learning algorithms seem to ignore these geometric considerations, while providing state of the art classification results on complex image data bases. It is thus time to wonder what type of image information is extracted by these deep neural networks and why do they work so well.

We show that deep convolution networks provide remarkable architectures to represent geometric image information, including shapes, textures and complex structured objects. They compute descriptors having appropriate invariance and stability to geometric transformations. The properties of these deep representations have strong similarities with Gestalt perception. However, the network filters do not need to be learned and can adapted to prior information on geometry, with multiscale wavelets. This brings us back from learning everything to more traditional computer vision approaches. Classification results will be shown on multiple image data bases, with linear SVM applied to network outputs.

Stéphane Mallat
École Normale Supérieure

Title: How Changing Mobile and Media Technologies is Changing The Way We Create Innovations

According to Schumpeter's definition of "Innovation," all the innovation instances are combinations of technologies that already exist. In that context, this talk covers the combination of progress of mobile network technologies and media understanding technologies. When the first mobile phones of 2nd Generation came out, people thought it was only for speech communication and texting. Then the phones got cameras, GPS. accelerometers, near filed communication devices together with fat communication pipes, people in all walks of life started using mobile phones in various opportunities. It is noteworthy that the progress of media understanding applications remarkably creates big data in a way of virtuous cycle by service and technology developments. This talk also highlights big-data driven service developments and API strategies for mobile innovations as well as technologies.

Minoru Etoh
NTT Docomo

Title: RGB-D Perception in Robotics

RGB-D cameras provide per pixel color and depth information at high frame rate and resolution. Gaming and entertainment applications such as the Microsoft Kinect system resulted in the mass production of RGB-D cameras at extremely low cost, also making them available for a wide range of robotics applications. In this talk, I will provide an overview of depth camera research done in the Robotics and State Estimation Lab over the last five years. This work includes 3D mapping, autonomous object modeling, unsupervised feature learning for object recognition, and articulated object tracking.

Dieter Fox
University of Washington