Programmable retinal dynamics in a CMOS mixed-signal array processor chip

The retina is responsible of the treatment of visual information at early stages. Visual stimuli generate patterns of activity that are transmitted through its layered structure up to the ganglion cells that interface it to the optical nerve. In this trip of micrometers, information is sustained by continuous signals that interact in excitatory and inhibitory ways. This low-level processing compresses the relevant information of the images to a manageable size. The behavior of the more external layers of the biological retina has been successfully modelled within the Cellular Neural Network framework. Interactions between cells are realized on a local basic. Each cell interacts with its nearest neighbors and every cell in the same layer follows the same interconnection pattern. Intra- and inter-layer interactions are continuous in magnitude and time. The evolution of the network can be described by a set of coupled nonlinear differential equations. A mixed-signal VLSI implementation of focal-plane low-level image processing based upon this biological model constitutes a feasible and cost effective alternative to conventional digital processing in real-time applications. A CMOS Programmable Array Processor prototype chip has been designed and fabricated in a standard technology. It has been successfully tested, validating the proposed design techniques. The integrated system consists of a network of 2 coupled layers, containing 32×32 elementary processors, running at different time constants. Involved image processing algorithms can be programmed on this chip by tuning the appropriate interconnection weights, internally coded as analog but programmed via a digital interface. Propagative, active wave phenomena and retina-lake effects can be observed in this chip. Low-level image processing tasks for early vision applications can be developed based on these high-order dynamics.

[1]  Ricardo Carmona-Galán,et al.  A VLSI-oriented continuous-time CNN model , 1996, Int. J. Circuit Theory Appl..

[2]  Tamás Roska,et al.  The CNN universal machine: an analogic array computer , 1993 .

[3]  A. Rodriguez-Vazquez,et al.  Four-quadrant one-transistor-synapse for high-density CNN implementations , 1998, 1998 Fifth IEEE International Workshop on Cellular Neural Networks and their Applications. Proceedings (Cat. No.98TH8359).

[4]  C. Rekeczky,et al.  Neuromorphic CNN models for spatio-temporal effects measured in the inner and outer retina of tiger salamander , 2000, Proceedings of the 2000 6th IEEE International Workshop on Cellular Neural Networks and their Applications (CNNA 2000) (Cat. No.00TH8509).

[5]  Leon O. Chua,et al.  The analogic cellular neural network as a bionic eye , 1995, Int. J. Circuit Theory Appl..

[6]  F. Werblin,et al.  Vertical interactions across ten parallel, stacked representations in the mammalian retina , 2001, Nature.

[7]  Carver Mead,et al.  Analog VLSI and neural systems , 1989 .

[8]  Frank Werblin,et al.  Synoptic Connections , Receptive Fields , and Patterns of Activity in the Tiger Salamander Retina , 2005 .

[9]  D. Balya,et al.  A qualitative model-framework for spatio-temporal effects in vertebrate retinas , 2000, Proceedings of the 2000 6th IEEE International Workshop on Cellular Neural Networks and their Applications (CNNA 2000) (Cat. No.00TH8509).

[10]  Tamás Roska,et al.  Methods for constructing physiologically motivated neuromorphic models in CNNs , 1996, International journal of circuit theory and applications.