29 October 2019 in IEEE Access
6 June 2017 in IEEE Internet of Things Journal
Deep neural network are one of the most powerful model for machine learning, which can learn the underlying patterns automatically from a large amount of data. So it can be extensively used in more and more Internet-of-Things (IoT) applications. However, the training of deep models is difficult, suffering from overfitting and gradient vanishing problem. Besides, the large amount of parameters and multiplication operations make it impractical for most deep learning models to directly execute on target hardware. In this paper, we propose a method of gradually pruning the weakly connected weights to improve the traditional stochastic gradient descent. And we adopt a reinforcement learning method called learning automata to find the weakly connected weights on account of its strong policy-making ability in stochastic and nonstationary environment. Our proposed method can learn a more effective and sparsely connected architecture during training from the initially fully connected neural networks. The experiments on MNIST show that our method have stronger power to defeat overfitting and can get better generalization performance on test set. Meanwhile, the thin and sparsely connected model we get can be more suitable for IoT applications.
26 March 2019 in IEEE Access
Deep learning methods have got fantastic performance on lots of large-scale datasets for machine learning tasks, such as visual recognition, neural language processing and so on. Most of the progresses on deep learning in recent years lied on supervised learning, for which the whole dataset with respect to a specific task should be well-prepared before training. However, in the real-world scenario, the labeled data associated with the assigned classes are always gathered incrementally over time, since it is cumbersome work to collect and annotate the training data manually. This suggests the manner of sequentially training on a series of datasets with gradually added training samples belonging to new classes, which is called incremental learning. In this paper, we proposed an effective incremental training method based on learning automata for deep neural networks. The main thought is to train a deep model with dynamic connections which can be either “activated” or “deactivated” on different datasets of the incremental training stages. Our proposed method can relieve the destruction of old features while learning new features for the new added training samples, which can lead to better training performance on the incremental learning stage. Experiments on MNIST and CIFAR-100 demonstrated that our method can be implemented for deep neural models in long sequence of incremental training stages and can achieve superior performance than training from scratch and the fine-tuning method.
19 November 2018 in IEEE Transactions on Emerging Topics in Computational Intelligence
Deep neural network has been one of the most powerful models in the field of machine learning, which has acquired state-of-the-art results in many tasks including image classification, object detection, text recognition, and so on. There have been many tricks to improve the training and generalization performance of deep neural network, such as dropout, ReLU, batch normalization, etc. In this paper, we proposed a new basic element to form deep neural networks, called learning automata competition unit (LCU). The LCU includes a group of general neural units and learning automata. The adopted learning automata are reinforcement learning methods, which can learn the optimal action through continuously interacting with a stochastic environment. Since the learning automata has powerful policy-making ability for both stochastic and non-stationary environment, the proposed LCU can facilitate competition in a group of neural units and gradually select the better trained neural units during training. The selected neural units through competition can make the training process more efficient, which can simultaneously get better training and generalization performance. The experiments on MNIST, CIFAR-10, and the Reuters newswire topic classification dataset show the performance of our method for both deep fully connected neural network and convolutional neural network.