Establishing the flow of information between two bio-inspired spiking neural networks

Abstract

In this article, four bio-inspired spiking neural networks (emitter1&2, receiver1&2) were presented. Using the proposed brain-compatible learning, spiking networks were trained with spike-based unsupervised weight optimization to recognize handwritten digits, capital letters, YALE, and ORL datasets. Recent experimental reports have proven that effective neuronal interactions are affected by neuronal synchronization. Therefore, by preserving synchronization between emitters and receivers, the effective connection between them was established. Then, transfer entropy was used to quantify the exchanged information between emitter1-receiver1 and emitter2-receiver2. We proved that depending on the synchronization metric between emitter and receiver, information of emitter could be transmitted to receiver network. With these prerequisites, receivers were fed via the transmitted information of emitters without training. The remarkable result was that receiver1 and receiver2 achieved 70.23% and 84.97% recognition accuracy, respectively, without being trained for the handwritten digits and YALE faces. This means that we hacked 78.38% of the information of emitter1 and transmitted them into receiver1 (also 88.79% of emitter2 information was transmitted to receiver2). Thus, receivers learned extra information without training, which was guided solely by the information provided by emitters. In this paper, for the first time, we were able to bypass the training process of the spiking networks.