Neuromorphic models of visual and motion processing in the fly visual system

Since the first neuromorphic retina was introduced 10 years ago, we have seen neuromorphic modeling extended to motion processing, saccadic systems, and auditory processing, to name a few. This dissertation extends neuromorphic modeling to the fly visual system. The retinotopic and regular arrangement of the layers in this system makes viable the mapping of the structure of the layers to silicon. The ability of the fly to compute motion reliably with only 24,000 receptors, while consuming only microwatts of power, also makes this system attractive for neuromorphic modeling. I start this dissertation by comparing the filter bandwidth properties and offsets of the fly receptors with those of silicon receptors. The filtering properties and the offsets of the receptors are critical because they determine the limits of subsequent processing circuitry. This work is the first characterization of biological and artificial offsets. Next, I describe an analog circuit that captures some of the adaptation and temporal filtering properties of the cells in the initial layers of the visual system. The adaptation time constant of the circuit is controllable via an external bias. The temporal filtering of the circuit changes with the S/N ratio of signals. This prefiltering preceding the motion areas is important to ensure that the motion computation is robust under different S/N rations. The filtering should be adaptive to match the S/N ratio so as to maximise the information transfer to subsequent processing. Adaptation is also a big component of the motion computation since the visual system has to extract information from a 2 to 3 decade range of speeds of objects under a six decade range of illumination. In this dissertation, I show the first adaptive motion model that matches its time scale to that of the moving image. The model explains experimental data showing motion adaptation in the direction-selective cells of the fly. Finally, I describe the responses of the direction-selective cells to motion. This silicon model is critical in showing that local direction selectivity can be computed from the correlation of continuous-time, graded inputs. The computation integrates the visual information over time in the decision making process and binarized features are not needed for the correlation. This model differs from previous silicon implementations of the Reichardt model that used token-based information for correlation. This model is a closer analogue of the motion computation in flies than previous silicon models