70 — Spiking Optical Flow for Event-based Sensors Using IBM’s TrueNorth Neurosynaptic System

Using IBM’s TrueNorth, a spiking neuron simulator framework living on bio-inspired hardware, Haessig et al present a technique for estimating optical flow (the anticipation and movement of a stimulus either over time or through space). Optical flow capability is common even in such simple brains as the fly’s, where it is used for high-speed 3D navigation and orientation during flight, but it remains elusively complex of a task in modern computer architectures, when accounting for power consumption and resource budgets. (ASIC and GPU implementations are quick, but very power inefficient.)

Using an ATIS — an “Asynchronous Time-based Image Sensor” that more closely approximates the data-collection nature of a retina — the authors combine the change detection intrinsic to that camera with simple neural networks taken from models of direction selective neurons in rabbit retina. These circuits are simulated on the TrueNorth hardware, where a simple time-dependent input race is used to excite or inhibit a DS unit.

Based on this “retina” circuitry, the authors are able to achieve direction-selective capability in a fully bio-inspired system using a fraction of the power required to drive a GPU- or ASIC-based model. Furthermore, the TrueNorth chip setup was only 57% utilized, which means that much more processing could take place on the chips than is currently done.

The authors suggest that this might be an opportunity for velocity estimation; but because velocity in a human-readable form is a continuous function, it would be required to use external processors to combine the signals to output a velocity float.