It was shown that it possible to acquire and follow a visual target with a fully neuromorphic close loop setup [1]. The system features a robot equipped with a Dynamic Vision Sensors (DVS) which extracts a target (blinking LED) from background activity. The relative position of the target is then stored in working memory (soft winner take all with stronger recurrent self-excitation). Due to background activity and low SNR in some situations the target is lost and can only be recovered with global information provided by a compass.
In order to overcome this problem with a neuromorphic solution, we could exploit other sensory modalities such as the Dynamic Audio Sensor (DAS). We propose to create a close loop neuromorphic auditory source following system consisting of:
- Dynamic Audio Sensor (DAS) [2]
- Spiking coincidence detector [3]
- Greedy Winner Take All (GWTA)
The sensory information is provided by a binaural (two ears) DAS (AMS1b / COCHLP) which has 64 frequency channels per ear and a temporal resolution of 1us.
The DAS is interfaced with cAER using a device with Linux OS (computer / Parallela / Raspberry Pi) in a such way that all the event can be collected. These events are send off to a Raggedstone which interfaces the Spiking coincidence detector implemented on a the latest test chip. The coincidence detector transforms an inter spike interval (ISI) between two standard Differential Pair Integrator (DPI) synapses into instantaneous firing frequency. We are going to use 3 coincidence detectors each consisting of 6 synapses and 2 output neurons (for positive and negative ISIs) [4]. This will give a range of [-200, +200] ms for the delay estimation between the events coming from the the two ears. These delays encode for the position of the target in the scene. In order to increase SNR and perform a decision where the desired target is relatively to the system, we propose to implement a greedy WTA which is characterized by a set of smaller soft winner take all with overlapping kernels. Furthermore, the normalization provided by the inhibitory population only act locally allowing multi target detection.
[1] Obstacle avoidance and target acquisition for robot navigation using a mixed signal analog/digital neuromorphic processing system
[2] Event-based 64-channel binaural silicon cochlea with Q enhancement mechanisms
[3] Neurally-inspired robotic controllers implemented on neuromorphic hardware
[4] Estimating the Location of a Sound Source with a Spike-Timing Localization Algorithm
