Abstract: The multiple hidden layers characteristic of deep neural networks was integrated into the framework of online sequential extreme learning machine. Thereby the author proposed a new algorithm called cost-sensitive regularized finite-memory multi-hidden-layer online sequential extreme learning machine. Cost sensitivity was embodied by the method of weighted least squares whereas the property of finite memory was embodied by discard of old obsolete data. Experiment results show that the online sequential extreme learning machines with the property of multiple hidden layers have higher image recognition accuracy and higher stability of recognition accuracy than that of single-hidden-layer.