Yenamachintala, Sai Sourabh (2018-12). FPGA Spiking Neural Processors With Supervised and Unsupervised Spike Timing Dependent Plasticity. Master's Thesis. Thesis uri icon

abstract

  • Energy efficient architectures for brain inspired computing have been an active area of research with recent advances in the field of neuroscience. Spiking neural networks (SNN) are a class of artificial neural networks in which information is encoded in discrete spike events, closely resembling the biological brain. Liquid State Machine (LSM) is a computational model developed in theoretical neuroscience to describe information processing in recurrent neural circuits and can be used to model recurrent SNNs. LSM is composed of an input, reservoir and output layers. A major challenge in SNNs is training the network with discrete spiking events for which traditional loss functions and optimization techniques cannot be applied directly. Spike Timing Dependent Plasticity (STDP) is an unsupervised learning algorithm which updates synaptic weights based on time difference between spikes of pre synaptic and post synaptic neurons. STDP is a localized learning algorithm and induces self-organizing behaviors resulting in sparse network structures making it a suitable choice for low cost hardware implementation. SNNs are hardware friendly as presence or absence of a spike can be encoded using a binary digit. In this research, SNN processor with energy efficient architecture is developed and is implemented on Xilinx Zynq ZC706 FPGA platform. Hardware friendly learning rules based on STDP are proposed and reservoir and readout layers are trained with these learning algorithms. In order to achieve energy efficiency, sparsification algorithm utilizing STDP rule is proposed and implemented. On chip training and inference are carried out and it is shown that with the proposed unsupervised STDP for reservoir training and supervised STDP for readout training, classification performance of 95% is achieved for TI corpus speech data set. Classification performance, hardware overhead and power consumption of the processor with different learning schemes are reported.
  • Energy efficient architectures for brain inspired computing have been an active area of research
    with recent advances in the field of neuroscience. Spiking neural networks (SNN) are a class of
    artificial neural networks in which information is encoded in discrete spike events, closely resembling
    the biological brain. Liquid State Machine (LSM) is a computational model developed in
    theoretical neuroscience to describe information processing in recurrent neural circuits and can be
    used to model recurrent SNNs. LSM is composed of an input, reservoir and output layers. A major
    challenge in SNNs is training the network with discrete spiking events for which traditional loss
    functions and optimization techniques cannot be applied directly. Spike Timing Dependent Plasticity
    (STDP) is an unsupervised learning algorithm which updates synaptic weights based on time
    difference between spikes of pre synaptic and post synaptic neurons. STDP is a localized learning
    algorithm and induces self-organizing behaviors resulting in sparse network structures making it
    a suitable choice for low cost hardware implementation. SNNs are hardware friendly as presence
    or absence of a spike can be encoded using a binary digit. In this research, SNN processor with
    energy efficient architecture is developed and is implemented on Xilinx Zynq ZC706 FPGA platform.
    Hardware friendly learning rules based on STDP are proposed and reservoir and readout
    layers are trained with these learning algorithms. In order to achieve energy efficiency, sparsification
    algorithm utilizing STDP rule is proposed and implemented. On chip training and inference
    are carried out and it is shown that with the proposed unsupervised STDP for reservoir training
    and supervised STDP for readout training, classification performance of 95% is achieved for TI
    corpus speech data set. Classification performance, hardware overhead and power consumption of
    the processor with different learning schemes are reported.

publication date

  • December 2018