A Survey of Feature Selection Methods in Deep Neural Networks – Deep learning, a technology which uses features as inputs to learn models, has been an open research area. Despite many attempts to use feature selection methods to make deep learning as a tool for machine learning, most of these work have focused on feature selection using two-part prediction or machine learning methods. While the two-part methods are successful for feature selection, they focus on the classification task and not on the real world. In this paper we propose a novel machine learning approach which combines the two-part prediction and classification processes to produce feature selections. The model can predict the feature set and the prediction process is conducted in a supervised fashion while learning the model. Our proposed algorithm uses a convolutional neural network to learn the classification task while the feature selection process is conducted in a supervised fashion. The proposed algorithm achieves an accuracy of 99.8% for the classification task and an accuracy of 99.8% for the real world task.
This paper presents a novel and effective learning approach for learning neural networks, which aims to obtain sparse representations of the input data (e.g., the neural network). This new approach consists of two key components. First, we first embed the input data into a sparse vector, based on its similarity between vectors. Our novel neural network is learned from the same learning task, without the need to directly classify the data. Next, a deep neural network is trained using the feature vectors extracted from the input data, which is then used to learn the network’s embedding. We evaluate our approach on the MNIST datasets, where it produces an error rate of 0.82 cm on average with a top-4 performance of 98.7% on CIFAR-10.
A new model of the central tendency towards drift in synapses
On the Impact of Lazy Recoding Theory on Bayesian Neural Networks
A Survey of Feature Selection Methods in Deep Neural Networks
Fast and reliable transfer of spatiotemporal patterns in deep neural networks using low-rank tensor partitioningThis paper presents a novel and effective learning approach for learning neural networks, which aims to obtain sparse representations of the input data (e.g., the neural network). This new approach consists of two key components. First, we first embed the input data into a sparse vector, based on its similarity between vectors. Our novel neural network is learned from the same learning task, without the need to directly classify the data. Next, a deep neural network is trained using the feature vectors extracted from the input data, which is then used to learn the network’s embedding. We evaluate our approach on the MNIST datasets, where it produces an error rate of 0.82 cm on average with a top-4 performance of 98.7% on CIFAR-10.
Leave a Reply