Sparse neural coding & bionic vision system
Brains are efficient low-power embedded machines that clearly outperform today's most powerful conventional computers at complex tasks such as vision or audition. Neuromorphic computing tries to understand the principles of brain processing by reverse engineering neural circuits and computational modeling to build systems that process information like biological neurons. However, bio-inspired systems need important amount of processing power and memory storage that can only be provided by modern High Performance Computers to useful computations to mimic the human cognition process. Unfortunately, the computing complexity of such algorithms makes their use on embedded systems often impossible without important simplifications. Hence, designing for example powerful embedded bio-inspired vision systems requires a new melding of engineering principles and technologies with knowledge gleaned from neuroscience.
The SENSE project will leverage on the recently proposed sparse neural model named GBNN (Gripon Berrou Neural Network). GBNN is a powerful, yet quite simple, neural network model that addresses the limitations of existing models like Hopfield or Boltzman networks in terms of efficiency and provides nearly optimal associative memories. GBNN being based on sparse coding, cluster and neural clique concepts is at the crossroads of distributed error correcting codes and graph theory. The neural networks that can be designed from these features offer wide perspectives in machine learning and artificial intelligence and are also very plausible from a biological point of view.
In this context, the SENSE project proposes to design radically different vision systems. Such tightly integrated and multi-layered smart vision systems will be organized into a stacked architecture that will use GBNN models at each level conferring unexpectedly powerful image processing. Image caption will be realized by designing a smart CMOS sensor embedding on-chip analog GBNN. Cortical processing and abstraction will be realized by designing a dedicated digital architecture implementing a hierarchical GBNN in order to further abstract visual information. Finally, in order to make the SENSE vision system tunable to target specific image processing applications, a software model of GBNN coupled with original image processing algorithms will be executed on a programmable many-core architecture. Besides, this work will allow not only to validate the GBNN model through a human vision system but also to project it on analog, digital and software technologies.
The consortium involves the following people and teams:
· UBS, Lab-STICC : Philippe Coussy, Cyrille Chavet, Laura Conde-Canencia
· Telecom-Bretagne, Lab-STICC : Sylvie Kerouedan, Cyril Lahuec
· UBS, Irisa : Sébastien Lefèvre, Nicolas Courty