Beyond Directed Acyclic Computation Graph with Cyclic Neural Network

This paper investigates a fundamental yet overlooked design principle of artificial neural networks (ANN): We do not need to build ANNs layer-by-layer sequentially to guarantee the Directed Acyclic Graph (DAG) property. Inspired by biological intelligence, where neurons form a complex, graph-structured network, we introduce the transformative Cyclic Neural Networks (Cyclic NN). It emulates biological neural systems' flexible and dynamic graph nature, allowing neuron connections in any graph-like structure, including cycles. This offers greater flexibility compared to the DAG structure of current ANNs. We further develop the Graph Over Multi-layer Perceptron, the first detailed model based on this new design paradigm. We experimentally validate the advantages of Cyclic NN on widely tested datasets in most generalized cases, demonstrating its superiority over current layer-by-layer DAG neural networks. With the support of Cyclic NN, the Forward-Forward training algorithm also firstly outperforms the current Back-Propagation algorithm.This research illustrates a transformative ANN design paradigm, a significant departure from current ANN designs, potentially leading to more biologically similar ANNs.