Без заголовка

--->> The SNAP spiking neural network learned the features of vehicles passing by the sensor within seconds. It detected and started counting cars in real time. The results of this hardware demonstration shows that SNAP can process events emitted by the DAVIS camera in real time and perform unsupervised learning of temporally correlated features. The DAVIS Dynamic Vision Sensor is an artificial retina that has an AER (Address Event Representation) interface, the same interface that is used in the SNAP Technology. The AER bus has become an industry standard. Rather than outputting frames of video, each pixel outputs one or more spikes whenever the contrast changes. A contrast change can be caused by movement. Spike events are transmitted over the AER bus at a maximum rate of 50 million events per second. The BrainChip SNAP technology can process 100 million events per second. The AVFE network autonomously learns to identify objects moving through its vision field. The second labelling neural network is trained to start counting these objects. The AFE system is comprised of a SNAP spiking neural network that learns from input patterns and performs autonomous feature extraction and data labeling for pattern recognition. The system can be used to process a wide range of real-world events and digital data from a multiple of sensors. The system can learn from both recorded and real-time input data. The learned features can be stored in a knowledge library. This library of learned behavior can be loaded on to further systems in order to instantaneously assimilate the learned features. The system is composed of three SNAP units. The first unit consists of sensory neurons that map input data to spike patterns that are distributed spatially and over time. Similar to a biological brain, these sensory neurons will fire spikes at different times depending on the input they receive. The outputs of the sensory neurons are then transmitted at a rate of up to 100 million events per second to the next unit that performs autonomous feature extraction, also called unsupervised feature learning. These 100 million events are distributed to thousands of synapses in parallel, which are updated one million times per second. This unit learns the main features that characterize a given set of data. For example, this unit learns the features of letters and digits when the input data consists of handwritten characters. The autonomous feature extraction unit is composed of a spiking neural network that utilizes an unsupervised learning rule (e.g. Spike Timing Dependent Plasticity: STDP), and lateral inhibition so that different neurons in the network learn different features. Lateral inhibition means that the first neuron to recognize a specific input pattern inhibits all others in the same layer. The learning process that occurs in this unit can run continuously or over a specific period of time. The SNAP learning method has proven to be very fast. The output of the autonomous feature extraction unit is forwarded to the labeling unit. This unit consists of a spiking neural network that is trained in a supervised way to map learned features to output labels. For the handwritten characters' example, the output labels could be letters and digits, or complete words that the system has learned to recognize. The output labels can be used in an external device like a Central Processing Unit (CPU) for post-processing. We look forward to providing further updates on the application of this exciting new technology through the year.