Han-Gyu Kim

Voice Command Recognition for Fighter Pilots Using Grammar Tree

This research copes with the voice command recognizer for fighter pilots. The voice command is composed of several connected words in the fighter system. And the recognizer automatically separates the command into individual words and implements isolated word recognition for each word. To improve the performance of the command recognizer, the error correction using grammar tree is proposed. The isolated word recognition error is corrected in the error correction process. Our experimental result shows that the grammar tree significantly improved the performance of the command recognizer.

Recurrent neural networks with missing information imputation for medical examination data prediction

In this work, we use recurrent neural network (RNN) to predict the medical examination data with missing parts. There often exist missing parts in medical examination data due to various human factors, for instance, because human subjects occasionally miss their annual examinations. Such missing parts make it hard to predict the future examination data by machines. Thus, imputation of the missing information is needed for accurate prediction of medical examination data. Among various types of RNNs, we choose simple recurrent network (SRN) and long short-term memory (LSTM) to predict the missing information as well as the future medical examination data, as they show good performance in many relevant applications. In our proposed method, the temporal trajectories of the medical examination measurements are modeled by RNNs with the missed measurements compensated, which is then used to predict the future measurements to be used as diagnosing the diseases of the subjects in advance. We have carried out experiments using a medical examination database of Korean people for 12 consecutive years with 13 medical fields. In this database, 11500 people took the medical check-up every year, and 7400 people missed their examination occasionally. We use complete data to train RNNs, and the data with missing parts are used to evaluate the imputation and future measurement prediction performance. In terms of root mean squared error (RMSE) and source to noise ratio (SNR) between the prediction and the actual measurements, the experimental results show that the proposed RNNs predicts medical examination data much better than the conventional linear regression in most of the examination items.

Medical examination data prediction using simple recurrent network and long short-term memory

In this work, we use two different types of recurrent neural networks (RNNs) to predict medical examination results of a subject given the previous measurements. The first one is a simple recurrent network (SRN) which models temporal trajectories of a data sequence to infer the unknown future observation, and the second one is a long short-term memory (LSTM) that enables modeling the longer trajectories by exploiting forgetting switches. The non-linear, temporal evolution of medical status of a human subjects are approximated by the RNNs, and the prediction of the future measurement becomes more accurate than those of the linear approximation method. The performance evaluation experiments are carried out on the real medical examination data, and the proposed methods show superior performances over the linear regression method. For the subjects who have abnormal behaviors in their medical examination results, the performance improvements are much more significant, so the proposed methods are expected to be used in detecting potential patients to provide earlier diagnosis and proper treatments for their illnesses.

Rehabilitation posture correction using deep neural network

The rehabilitation treatment is important because it helps a patient restore physical sensory and mental capabilities. The patient whose symptoms are moderately relieved, or outpatient, usually rehabilitate the individual alone. Improper exercise or posture can slow the recovery of the patient or even worsen the patients health status when doing rehabilitation exercise alone. The best way is to receive home visiting treatment from professional therapist until cured. However, such way is a burden on the patient in terms of cost. This paper proposes the novel model that corrects the improper postures of the patient when having rehabilitating exercise alone. We use Microsoft Kinect to recognize the posture of the patient by extracting the human skeleton. We will adopt deep neural network to analyze the extracted human skeleton, in order to determine whether the posture is correct or not. The data for training our model will be correct postures and incorrect postures and detailed data collection plan is provided in this paper. The implementation and experiment will be performed in the future work.

Controlled dropout: A different approach to using dropout on deep neural network

Deep neural networks (DNNs), which show outstanding performance in various areas, consume considerable amounts of memory and time during training. Our research led us to propose a controlled dropout technique with the potential of reducing the memory space and training time of DNNs. Dropout is a popular algorithm that solves the overfitting problem of DNNs by randomly dropping units in the training process. The proposed controlled dropout intentionally chooses which units to drop compared to conventional dropout, thereby possibly facilitating a reduction in training time and memory usage. In this paper, we focus on validating whether controlled dropout can replace the traditional dropout technique to enable us to further our research aimed at improving the training speed and memory efficiency. A performance comparison between controlled dropout and traditional dropout is carried out by implementing an image classification experiment on data comprising handwritten digits from the MNIST dataset (Mixed National Institute of Standards and Technology dataset). The experimental results show that the proposed controlled dropout is as effective as traditional dropout. Furthermore, the experimental result implies that controlled dropout is more efficient when an appropriate dropout rate and number of hidden layers are used.

A Novel Concept of the Rehabilitation Training Coach Robot for Patients with Disability

This paper proposes the rehabilitation treatment coach robot which will help at-home patients do their rehabilitation exercises at home without any professional trainers. The coach robot is designed to be cheap enough for patients to afford it. The robot suggests the rehabilitation program and corrects the posture of the patients during the exercise. The deep neural network is used for posture correction. Besides, the voice interface is applied for convenient interaction between robot and patients during the exercise. The emergency detection module is adopted which will inform doctors when emergency happens on patients. The emergency detection will be implemented using deep neural network on voice input and video input simultaneously. The detailed data collection plan for training deep neural network and performance evaluation plan are also provided in the paper.

Model Regularization of Deep Neural Networks for Robust Clinical Opinions Generation from General Blood Test Results

The deep neural network (DNN) that models characteristics of general blood test (GBT) results was used in clinical opinions generation. The DNN that generates clinical opinions has the complex structure, which causes overfitting problem. The relatively small size of medical dataset also contributes to the occurrence of overfitting. In order to deal with overfitting, we apply two techniques that solve overfitting of DNN, which are dropout, and batch normalization. Dropout is inserted into the network in various ways in order to find out the optimal structure of the network. Batch normalization is also added in various ways for the same purpose. The experiment conducted on GBT dataset shows that DNNs with dropout and batch normalization outperform the simple DNN in generating clinical opinions for our GBT dataset. Besides, dropout shows slightly better performance compared to batch normalization.

Clinical opinions generation from general blood test results using deep neural network with principle component analysis and regularization

The conventional approach of generating clinical opinions from general blood test (GBT) results uses the deep neural network (DNN) comprised of fully-connected layers. The large number of input neurons and output neurons result in the complex DNN structure, which causes overfitting problem. However, the dimension of the input vector and the output vector cannot be reduced arbitrarily, as all GBT results and all clinical opinions should be retained. In order to avoid overfitting, we apply principal component analysis (PCA) and parameter regularization. PCA is a dimensionality reduction technique which may be used to reduce the number of input neurons, minimizing the information loss. Besides, we apply L1 penalty or L2 penalty to the loss function of the DNN to apply parameter regularization. We also apply PCA and the regularization simultaneously. Experimental results show that all three proposed methods outperform the conventional DNN, and applying only L1-regularization shows the best performance in avoiding overfitting in the DNN for generating clinical opinions.

Discriminative restricted Boltzmann machine for emergency detection on healthcare robot

In this work, we propose a concept of emergency detection algorithm for healthcare robot which adopts discriminative restricted Boltzmann machine for anomaly detection. We will adopt anomaly detection rather than simple emergency case classification as it is hard to collect real emergency data to train the effective classifier. The conventional anomaly detection method uses decision tree to analyze the signals obtained from the sensors attached on the bodies of the patients to find out the emergency situations. We propose anomaly detection using video and audio signals as they are easy to be obtained by the healthcare robot, with equipping a camera and a microphone, and it is much more convenient for patients. The discriminative restricted Boltzmann machine which is specialized in learning probability distribution in an unsupervised manner will be applied for anomaly detection. This paper only provides the novel idea for emergency detection. The implementation and the experiments will be conducted in the future work.