Unsupervised shape learning in a neuromorphic hierarchy

We present a neural-based learning system for object recognition in still gray-scale images. The system comprises several hierarchical levels of increasing complexity modeling the feed-forward path of the ventral stream in the visual cortex. It learns typical shape patterns of objects as these appear in images from experience alone without any prior labeling. Information about the exact origin of parts of the stimulus is systematically discarded, while the shape-related object identity information is preserved, resulting in strong compression of the original image data. To demonstrate it’s capabilities, we train the system on publicly available image databases and use it’s final output in classification tasks. The text was submitted by the authors in English. Daniel Oberhoff received a diploma in physics in 2004 from ETH Zurich, where he spent over a year in the Physics of New Materials group of the solid state physics department working on the characterization and simulation of organic electronic devices. In the fall of 2005, he began work on his PhD thesis on biologically inspired machine learning for computer vision with Fraunhofer Gesellschaft at Schloss Birlinghoven in Sankt-Augustin, Germany. Marina Kolesnik received her masters degree in physics and engineering in 1984 from Moscow Institute of Physics and Technology and a PhD in computer science in 1993 from the Russian Academy of Sciences. From 1984 to 1995, she worked as a research scientist at the Space Research Institute in Moscow, Russia. In 1995–1996, she was a visiting researcher as a Lise-Meitner fellow at JOANNEUM Research in Graz, Austria. She obtained a Max-Planck fellowship in 1996. Since 1997, she has been a research scientist at the GMD Research Center and, since 2002, at Fraunhofer Gesellschaft in Sankt-Augustin, Germany. Her research interests include computer vision, vision-based operations, and visual perception and modeling of visual pathways for computer vision applications.