Exploiting the Sparsity of Deep Neural Networks for Predictive-Advection Mining – This paper presents a new technique to efficiently and efficiently process a Convolutional Neural Network (CNN), while keeping the network stable. After several hours, CNNs are being trained independently in an online fashion, which allows us to effectively improve the performance of the CNN in a supervised fashion. We implement this idea into a novel method for fast learning using ImageNet, and analyze its performance using a well-validated deep CNN. Results show that our algorithm can improve the CNN for the classification task, while maintaining the stability of the network.
This study proposes a new technique for 3D reconstruction from partial deformation measurements in low- and high-resolution datasets. This is accomplished by constructing the partial measurements for each point in a deformable space, based on a mapping scheme of a data-rich optical flow. This mapping scheme is also exploited by extracting the high-resolution reconstruction from data. To tackle the problem of high-resolution deformation measurements, the proposed technique is first applied to a large dataset of deformable signals, and then combines the reconstructed partial measurements to improve the reconstruction performance. Experiments on simulated and real deformation measurements indicate that the proposed approach achieves comparable results compared to state-of-the-art methods.
Online Model Interpretability in Machine Learning Applications
Practical algorithms, networks and neural nets
Exploiting the Sparsity of Deep Neural Networks for Predictive-Advection Mining
Unsupervised learning with spatial adversarial filtering
Robust Depth Map Estimation Using Motion Vector RepresentationsThis study proposes a new technique for 3D reconstruction from partial deformation measurements in low- and high-resolution datasets. This is accomplished by constructing the partial measurements for each point in a deformable space, based on a mapping scheme of a data-rich optical flow. This mapping scheme is also exploited by extracting the high-resolution reconstruction from data. To tackle the problem of high-resolution deformation measurements, the proposed technique is first applied to a large dataset of deformable signals, and then combines the reconstructed partial measurements to improve the reconstruction performance. Experiments on simulated and real deformation measurements indicate that the proposed approach achieves comparable results compared to state-of-the-art methods.