Kategorie: IS-Colloquium

Abstract: I will describe work in my group over the last 2 years developing a theory of visual cortex and of deep learning architectures of the convolutional type. I will describe the theoretical consequences of a simple assumption: the main computational goal of the feedforward path in the ventral stream – from V1, V2, V4 and to IT – is to discount image transformations, after learning them during development. The initial assumption is that a basic neural operation consists of dot products between input vectors and synaptic weights – which can be modified by learning. I will outline theorems showing that a multi-layer hierarchical architecture of dot-product modules can learn in an unsupervised way geometric transformations of images and then achieve the dual goals of invariance to global affine transformations and of robustness to deformations. These architectures develop to be automatically invariant to transformations of a new object, achieving the goal of recognition with very few labeled examples. The theory should apply to a range of hierarchical architectures such as HMAX, convolutional “deep learning” networks and related feedforward models of sensory systems and extend their power.
[mehr]

Kategorie: Öffentliche Veranstaltungen

The focus of this introduction is CUDA, the parallel computing platform created by NVIDIA, and the OpenACC programming standard. During the workshop, professional NVIDIA trainers will give talks on both topics.
[mehr]

Kategorie: FriKo

Speaker: Anna-Marie Brouwer, TNO Perceptual and Cognitive Systems
[mehr]

Kategorie: MoKo/MiKo

Prof. Udo Heinemann, Max Delbrück Zentrum Berlin-Buch Abstract: TBA
[mehr]

Kategorie: FriKo

Speaker: TBA
[mehr]