Tobi Delbruck PhD. Thesis
This thesis is a detailed description of a neuromorphic visual-motion processing
chip and its component parts. The chip is the first two-dimensional silicon
retina with a full set of direction-selective, velocity-tuned pixels. The
architecture for the chip is based on the biological correlation-type motion
detector, with the addition of a novel spatiotemporal aggregation. All the
processing on the chip is analog and occurs in parallel. Novel, on-chip,
continuous-time, adaptive, logarithmic photoreceptor circuits are used to
couple temporal image signals into the motion processing network. These
continuous-time photoreceptor circuits have also been used in a wide variety
of other vision chips. The photoreceptor circuits center their operating
point around the history of the illumination, simultaneously achieving high
sensitivity and wide dynamic range. The receptor circuits are characterized
and analyzed carefully for their temporal bandwidth and detection performance.
Noise properties are analyzed, resulting in a simple and intuitive understanding
of the limiting parameters. Novel adaptive elements are described that are
insensitive to light-generated minority carriers. Novel measurements are
presented of the spectral response properties of phototransducers that can
be built in ordinary CMOS or BiCMOS processes. A novel nonlinear circuit
that measures similarity and dissimilarity of signals is described and characterized.
These bump circuits are used on the motion chip to extract the motion energy
signal, and have also been used in other chips in numerous ways.
September 13, 2007