We expect that CAVIAR will be a significant milestone for vision research. The most powerful vision algorithms, which are those based on bio-inspired inter-layer massive connectivity using perceptive fields or, equivalently, convolution-type processing, cannot be implemented with conventional hardware for real-time processing. CAVIAR wants to develop and make available a set of hardware elements and infrastructure that can provide computational neuroscience researchers with an invaluable tool to mimic biological vision systems in real time.
CAVIAR results are
not only limited to vision research, but can be extended to other neuronal
disciplines. The availability of general purpose programmable-kernels low-level convolutional layers and higher level processing and learning layers that work in real time, can be exploited to mimic more generic neural structures that require adaptation as well.
With the development
of a flexible programmable AER infrastructure, CAVIAR will provide a medium
investigating current and novel models of spike-based processing in systems that interact with a real environment. Both rate coding and temporal coding ideas can be tested especially in regards with two components in CAVIAR: The "object" chip and the spatio-temporal learning chip both receive AER inputs and produce AER outputs. The basic computing units on these two chips comprise spiking neurons and synapses. On these chips, temporal coding ideas will be explored within this proposal. We expect to gain new insights and ideas on properties of temporal coding when implemented on a physical system that interacts with a real environemnt.
CAVIAR is conceived in principle to enable a longer term research towards application in “Automatic Vehicle Driving”. Nevertheless, CAVIAR intends to demonstrate its potential by developing an appropriate “demonstrator” at the end of the four years work. The objective is to use all the elements developed during the CAVIAR project to assemble a small multi-layer vision system able to recognize a ball-shaped object, independently of lighting conditions and background. Such system will be mounted on a small commercial robot, enabling it to detect and follow a ball rolling at high speeds.
CAVIAR can clearly
be classified into ‘Neuromorphic Engineering’ research for ‘Perception’
However, its results can be extended to any kind of sensing mechanism (vision, audition, olfaction, sonar, ...). It will provide an invaluable tool for researchers in the ‘Cognitive/Computational Neuroscience’ field or to industries willing to exploit compact and real-time vision (or other bio-like) perception systems.