Frequency-Dependent Synaptic Plasticity Model for Neurocomputing Applications

In neuroscience, there is substantial evidence that suggests temporal filtering of stimulus by synaptic connections. In this paper, a novel frequency-dependent plasticity mechanism (FDSP) for neurocomputing applications is presented. It is proposed that synaptic junctions could be used to perform bandpass filtering on the input stimulus. The unique transfer function of a bandpass filter replaces the conventional weight value associated with synaptic connections. The proposed model has been simulated and rigorously tested with standard machine learning benchmarks such as XOR and multivariate IRIS dataset while utilising minimum resources. The proposed model offers a unique advantage and has the potential to overcome the burden of hidden layer neurons from the network. Exclusion of hidden layer from the network significantly reduces the size of the network and hence the computational effort required for classification tasks. The proposed FDSP mechanism allows for complete analogue system design with a frequency multiplexed communication scheme. The main goal of this study is to establish frequency-dependent plasticity as an alternative to existing time-domain based techniques. The proposed method has a number of applications in neurocomputing, low power IoT devices and compute-efficient Deep Convolutional Neural Networks (DCNNs).