AN ADVANCED FRAMEWORK TO OPTIMIZE MULTI-RESIDUAL RECURRENT NEURAL NETWORK FOR BETTER SEQUENCE LEARNING
Loading...
Date
2019-11-22
Authors
Journal Title
Journal ISSN
Volume Title
Publisher
Abstract
Sequence Learning is the cornerstone of data mining, and is significant in extracting useful information, from sequencing sounds in a speech to sequencing semantics in linguistics. Before sequence learning, finding a proper approach to collect sequential data is essential. The main purpose of the proposed recognition system is to defeat the CAPTCHA (Completely Automated Public Turing test to Tell Computers and Humans Apart) because we need to collect the data for the sequence learning. Besides, defeating the CAPTCHAs is also beneficial to improving the safety when we expose the CAPTCHAs’ deficiency. As an effective way to protect the security and preserve the privacy of the network data, CAPTCHA is widely used in recent years. Normally, three steps are utilized to defeat the CHAPCHAs - Preprocessing, Segmentation and Recognition. Since there is not a universal segmentation framework that is adaptive to all the possible CAPTCHA characters, each individual character requires separate segmentation which makes the segmentation complicated. In this dissertation, we present a self-adaptive algorithm in optimally segmenting different CAPTCHA characters. Current classifiers including Template Matching (TM), Optical Character Recognition (OCR) and Convolutional Neural Networks (CNN) are utilized in classifying these segmented CAPTCHA characters. The CAPTCHAs experimental results show the outperformance of the proposed recognition system in defeating the CAPTCHA. In the currently existing financial related Chinese text classification task, the data quality of those tasks is not ideal because labeled Chinese datasets are not large enough.
Besides the textbased CAPTCHA recognition, short-term text classification also plays an important role in sequence learning. After obtaining the titles of Chinese commercial news, a new Chinese financial related Short-term Text Classification Task (STCT) is introduced and its corresponding benchmark is provided. As a popular solution for STCT in sequence learning, recurrent neural networks (RNNs) have proven its efficiency in processing sequential information. However, the traditional RNNs have suffered from the gradient diminishing problem until the advent of Long Short-Term Memory (LSTM). The LSTM, though, is still weak in capturing long-time dependency in sequential data due to the inadequacy of memory capacity in LSTM cells. To address this challenge, we propose an Attention-augmentation Bidirectional Multi-residual Recurrent Neural Network (ABMRNN) to overcome this deficiency. The proposed ABMRNN integrates both past and future information at every time step with an omniscient attention model. The multi-residual mechanism has also been proposed in our model targeting the pattern of the relationship between the current time step and further distant time steps instead of only one previous time step. The experimental results show that the proposed model outperforms the traditional statistical classifiers and other state-of-theart variations of RNN architectures in both the STCT and other public tasks, such as AG news, Sequential-MNIST, and IMDB.
Description
Keywords
sequence learning, recurrent neural network, residual network, multi-residual, CAPTCHA