Bidirectional stochastic configuration network for regression problems

To adapt to the reality of limited computing resources of various terminal devices in industrial applications, a randomized neural network called stochastic configuration network (SCN), which can conduct effective training without GPU, was proposed. SCN uses a supervisory random mechanism to assign its input weights and hidden biases, which makes it more stable than other randomized algorithms but also leads to time-consuming model training. To alleviate this problem, we propose a novel bidirectional SCN algorithm (BSCN) in this paper, which divides the way of adding hidden nodes into two modes: forward learning and backward learning. In the forward learning mode, BSCN still uses the supervisory mechanism to configure the parameters of the newly added nodes, which is the same as SCN. In the backward learning mode, BSCN calculates the parameters at one time based on the residual error feedback of the current model. The two learning modes are performed iteratively until the prediction error of the model reaches an acceptable level or the number of hidden nodes reaches its maximum value. This semi-random learning mechanism greatly speeds up the training efficiency of the BSCN model and significantly improves the quality of the hidden nodes. Extensive experiments on ten benchmark regression problems, two real-life air pollution prediction problems, and a classical image processing problem show that BSCN can achieve faster training speed, higher stability, and better generalization ability than SCN. (C) 2021 Elsevier Ltd. All rights reserved.

Automated summary

A novel bidirectional SCN algorithm (BSCN) is proposed to accelerate training speed and improve the quality of hidden nodes, achieving faster speed, higher stability, and better generalization ability than SCN in various experiments.