Jian (Skyler) Zheng

Classifying severely occluded images is a challenging yet highly- needed task. In this paper, motivated by the fact that human being can exploit context information to assist learning, we apply convolutional recurrent neural network (CRNN) to attack this challenging problem. A CRNN architecture that integrates convolutional neural network (CNN) with long short-term memory (LSTM) is presented. Three new datasets with severely occluded images and con- text information are created. Extensive… Show more

Classifying severely occluded images is a challenging yet highly- needed task. In this paper, motivated by the fact that human being can exploit context information to assist learning, we apply convolutional recurrent neural network (CRNN) to attack this challenging problem. A CRNN architecture that integrates convolutional neural network (CNN) with long short-term memory (LSTM) is presented. Three new datasets with severely occluded images and con- text information are created. Extensive experiments are conducted to compare the performance of CRNN against conventional methods and human experimenters. The experiment results show that the CRNN outperforms both conventional methods and most of the human experimenters. This demonstrates that CRNN can effectively learn and exploit the unspecified context information among image sequences, and thus can be an effective approach to resolve the challenging problem of classifying severely occluded images. Show less

In this paper we propose a novel network module, namely Robust Attentional Pooling (RAP),
that potentially can be applied in an arbitrary network for generating single vector representations
for classification. By taking a feature matrix for each data sample as the input,
our RAP learns data-dependent weights that are used to generate a vector through linear
transformations of the feature matrix. We utilize feature selection to control the sparsity in
weights for compressing the… Show more

In this paper we propose a novel network module, namely Robust Attentional Pooling (RAP),
that potentially can be applied in an arbitrary network for generating single vector representations
for classification. By taking a feature matrix for each data sample as the input,
our RAP learns data-dependent weights that are used to generate a vector through linear
transformations of the feature matrix. We utilize feature selection to control the sparsity in
weights for compressing the data matrices as well as enhancing the robustness of attentional
pooling. As exemplary applications, we plug RAP into PointNet and ResNet for point cloud
and image recognition, respectively. We demonstrate that our RAP significantly improves
the recognition performance for both networks whenever sparsity is high. For instance, in
extreme cases where only one feature per matrix is selected for recognition, RAP achieves
more than 60% improvement over PointNet in terms of accuracy on the ModelNet40 dataset. Show less

In this paper, we develop a novel scheme to reduce the training
data requirement in deep neural networks (DNNs). We
first apply a partial mutual information (PMI) technique to
seek for the optimal feature set for the DNN. Then we use
a correlation matching based active learning (CMAL) technique
to select and label the most informative training data.
We integrate these techniques in a DNN consisting of two layers
of unsupervised sparse autoencoders for feature… Show more

In this paper, we develop a novel scheme to reduce the training
data requirement in deep neural networks (DNNs). We
first apply a partial mutual information (PMI) technique to
seek for the optimal feature set for the DNN. Then we use
a correlation matching based active learning (CMAL) technique
to select and label the most informative training data.
We integrate these techniques in a DNN consisting of two layers
of unsupervised sparse autoencoders for feature extraction
and a supervised softmax layer for classification. Simulations
are conducted over the breast cancer data set from the UCI
repository to show that this scheme can drastically reduce the
labeled data needed for DNN training, and superior performance
of the DNN classification can still be achieved with a
reduced training data set. Show less

In this paper, a new active learning scheme is proposed for linear
regression problems with the objective of resolving the insufficient
training data problem and the unreliable training data labeling problem.
A pool-based active regression technique is applied to select the
optimal training data to label from the overall data pool. Then, compressive
sensing is exploited to remove labeling errors if the errors
are sparse and have large enough magnitudes, which are called… Show more

In this paper, a new active learning scheme is proposed for linear
regression problems with the objective of resolving the insufficient
training data problem and the unreliable training data labeling problem.
A pool-based active regression technique is applied to select the
optimal training data to label from the overall data pool. Then, compressive
sensing is exploited to remove labeling errors if the errors
are sparse and have large enough magnitudes, which are called large
outliers. Next, in order to mitigate the non-sparse labeling errors that
have relatively small magnitudes, which are called small outliers, a
new technique is developed to convert them back into sparse large
outliers. With both artificial and real data sets, extensive simulations
are conducted to verify the robustness of the proposed scheme in
training data selection and outlier suppression. Show less

In this letter, we develop an active learning algorithm to optimize the selection of training data for robust linear regression. This algorithm selects training data based on the principle of correlation matching between the training dataset and the overall data pool. Considering the inevitable and potentially heavy human labeling errors, we model the probability of labeling errors based on the item response theory (IRT) and develop data screening techniques to control the error sparsity… Show more

In this letter, we develop an active learning algorithm to optimize the selection of training data for robust linear regression. This algorithm selects training data based on the principle of correlation matching between the training dataset and the overall data pool. Considering the inevitable and potentially heavy human labeling errors, we model the probability of labeling errors based on the item response theory (IRT) and develop data screening techniques to control the error sparsity. Compressive sensing theory is then exploited for human labeling error suppression. This algorithm is robust even in the case of short training dataset with nonsparse labeling errors. Its performance is verified by simulations with both artificial data and real benchmark data. Experiments are also conducted to demonstrate the validity of the IRT-based human labeling error model and the superior performance of the algorithm in practical applications. Show less

In this paper, we propose a joint machine learning and human learning design approach to make the training data labeling task in linear regression problems more efficient and robust to noise, modeling mismatch, and human labeling errors.Considering a sequential active learning scheme which relies on human learning to enlarge training data set, we integrate it with sparse outlier detection algorithms to mitigate the inevitable human errors during training data labeling. First, we assume… Show more

In this paper, we propose a joint machine learning and human learning design approach to make the training data labeling task in linear regression problems more efficient and robust to noise, modeling mismatch, and human labeling errors.Considering a sequential active learning scheme which relies on human learning to enlarge training data set, we integrate it with sparse outlier detection algorithms to mitigate the inevitable human errors during training data labeling. First, we assume sparse human errors and formulate the outlier detection as a sparse optimization problem within the sequential active learning procedure. Then, for non-sparse human errors, with the IRT(item response theory) to model the distribution of human errors, appropriate data are selected to reconstruct a training data set with sparse human errors. Simulations are conducted to verify the desirable performance of the proposed approach.
Show less

In this paper, we propose a spectrum sharing
technique based on the compressive sensing theory. With this
technique, machine-to-machine (M2M) communication devices
conduct transmission using the same spectrum and channel with
the primary users such as the 5G cellular device. The known
and repetitive symbols of the M2M packets are exploited to
create the sparsity. Then compressive sensing algorithm is used
to detect jointly the M2M packets and the primary user’s… Show more

In this paper, we propose a spectrum sharing
technique based on the compressive sensing theory. With this
technique, machine-to-machine (M2M) communication devices
conduct transmission using the same spectrum and channel with
the primary users such as the 5G cellular device. The known
and repetitive symbols of the M2M packets are exploited to
create the sparsity. Then compressive sensing algorithm is used
to detect jointly the M2M packets and the primary user’s packet.
This technique emulates human’s attention capability in situations
such as the cocktail party problem. The performance is analyzed,
simulated, and demonstrated in a wireless transmission testbed. Show less

Electric load forecasting plays a vital role in smart grid. Short term electric load forecasting forecasts the load that is several hours to several weeks ahead. Due to the nonlinear, non-stationary and nonseasonal nature of the electric load time series, accurate forecasting is challenging. This paper explores Long-Short-Term-Memory (LSTM) based Recurrent Neural Network (RNN) to deal with this challenge. LSTM-based RNN is able to exploit the long term dependencies in the electric load time… Show more

Electric load forecasting plays a vital role in smart grid. Short term electric load forecasting forecasts the load that is several hours to several weeks ahead. Due to the nonlinear, non-stationary and nonseasonal nature of the electric load time series, accurate forecasting is challenging. This paper explores Long-Short-Term-Memory (LSTM) based Recurrent Neural Network (RNN) to deal with this challenge. LSTM-based RNN is able to exploit the long term dependencies in the electric load time series for more accurate forecasting. Experiments are conducted to demonstrate that LSTM-based RNN is capable of forecasting accurately the complex electric load time series with a long forecasting horizon. Its performance compares favorably to many other forecasting methods. Show less