pod3275/real-world-deep-learning-application-problems

Methods to Overcome the Real-World Deep Learning Application Problems

Methods to Overcome the Real-World Deep Learning Application Problems

ABSTRACT

In recent years, the development of machine learning, especially deep learning, and its applications have been actively studied. Although deep learning is applied in many areas, there are still problems to be solved in order to introduce deep learning models into real-world systems. In this paper, we summarized the papers that overcome various problems in the real-world application of the deep learning model.

1. Detect Adversarial Example Using Gaussian Process-based Detector

For the image classification model, an adversarial attack has been proposed which causes malfunction of the model by mixing noise that is hardly distinguishable by human into the input image of the model. As a result, the reliability of the deep learning model has been raised, and a study has been proposed to prevent the adversarial attack. Lee et al. proposed an adversarial detection method based on gaussian process regression using intermediate features extracted from the classification model. The proposed detector showed higher detection performance than other detection methods when extremely few adversarial examples are used in training.

2. String CAPTCHA attack using CNN

Lee et al. proposed a deep learning model that recognizes string CAPTCHAs that widely used in internet sites. The proposed method eliminates the noise in the CAPTCHA image through image processing, separates string image into single character images, and recognizes CAPTCHA characters by training CNN model. As a result of experiment on CAPTCHA used in Korea ticket reservation site, the proposed model showed a high recognition rate of 85% based on CAPTCHA.

3. Box Office Forecasting using Non-Linear Regression

Won et al. proposed a deep learning model that predicts movie audience demand before opening by using a nonlinear regression model. In addition to features provided by the KOFIC, which provides cinema information, they define features that can be used for prediction and proposed a Bi-LSTM deep learning model to perform sentimental analysis on movie reviews. Experimental results showed that the proposed method had higher performance than other sentimental analysis methods and effectively predicted the demand of movie audience.

4. Embedding for Out of Vocabulary Words Considering Contextual and Morphosyntactic Information

Won et al. proposed a technique for effectively handling the OOV words that are not in the existing vocabulary during word embedding. The proposed method is a Bi-LSTM structured deep learning model named Context-Char, which embeds the OOV words using contextual and morphosyntactic information. As a result of experiments using datasets containing the OOV words, the proposed method showed higher performance than other methods of handling OOV words.

5. Deep representation model-based Under-sampling Method for Imbalanced data

Lim et al. proposed an effective under-sampling technique for processing imbalanced data. They proposed deep representation models that extract the structural features of major class data and minor class data and calculated the degree of conformity of the data to determine under-sampling. In the experiments using various imbalanced data, the proposed method showed higher performance than other processing methods for most datasets.