Artificial Intelligence and Machine Learning algorithms help humans in various applications. Neural Networks systems are one of this area’s most important research topics, inspired by the human brain. In this field, Spiking Neural Networks (SNN) use spikes to communicate between neurons mimicking the brain’s algorithm. The input data produced by sensors has to be converted to spikes for training and testing these systems. Rate encoding is a popular method that encodes the input signal into the spiking frequency.
This work presents two methods to design an analog input encoder that receives the information and converts them to spiking output. Both ways use a ∆Σ modulator to create a digital output from the input signal. The first input encoder, called synchronous ∆Σ analog to spike converter, reads the digital output of the ∆Σ and produces a spike for every ’1’ bit. The second design is called neuromorphic ∆Σ analog to spike converter, which uses a synapse and neuron model to produce the rate-encoded spiking output. The synapse converts the ∆Σ output to a current, and the neuron receives this current at its input.
This thesis is the first design to build a general input encoder that can be used in most SNN systems. A clock signal can change the firing frequency of both encoders. The synchronous ∆Σ A-S converter can perform for clock signals between 1 kHz and 4 MHz, while the neuromorphic one can perform between 1 kHz and 2 MHz. The optimized clock frequency is 50 kHz for both of them. With this clock, the synchronous one’s accuracy is 99.2% encoding a DC input, and its input can have a maximum bandwidth of 120 Hz to achieve an SNR higher than 50 dB. It consumes 13.4 μW average power with 500 μm2 area. The neuromorphic one’s accuracy for DC inputs is 97.3%, and its maximum bandwidth is 65 Hz. It consumes 12.7 μW average power with 0.011 mm2 area.