• The Transactions of the Korean Institute of Electrical Engineers D
• /
• v.54 no.7
• /
• pp.468-474
• /
• 2005

Brain-machine interface (BMI) based on neuronal spike trains is regarded as one of the most promising means to restore basic body functions of severely paralyzed patients. The spike train decoding algorithm, which extracts underlying information of neuronal signals, is essential for the BMI. Previous studies report that a linear filter is effective for this purpose and there is no noteworthy gain from the use of nonlinear mapping algorithms, in spite of the fact that neuronal encoding process is obviously nonlinear. We designed several decoding algorithms based on the linear filter, and two nonlinear mapping algorithms using multilayer perceptron (MLP) and support vector machine regression (SVR), and show that the nonlinear algorithms are superior in general. The MLP often showed unsatisfactory performance especially when it is carelessly trained. The nonlinear SVR showed the highest performance. This may be due to the superiority of the SVR in training and generalization. The advantage of using nonlinear algorithms were more profound for the cases when there are false-positive/negative errors in spike trains.

• Proceedings of the Korea Information Processing Society Conference
• /
• 2022.11a
• /
• pp.679-679
• /
• 2022

• Journal of Biomedical Engineering Research
• /
• v.28 no.5
• /
• pp.694-702
• /
• 2007

The adaptation effect originating from the chemical synapse between auditory nerve and inner hair cell gives advantage in accurate representation of temporal cues of incoming speech such as speech onset. Thus it is expected that the modification of conventional speech processing strategies of cochlear implant(CI) by incorporating the adaptation effect will result in considerable improvement of speech perception performance such as consonant perception score. Our purpose in this paper was to evaluate our new CI speech processing strategy incorporating the adaptation effect by the observation of auditory nerve responses. By classifying the presence or absence of speech from the auditory nerve responses, i. e. spike trains, we could quantitatively compare speech onset detection performances of conventional and improved strategies. We could verify the effectiveness of the adaptation effect in improving the speech onset representation characteristics.

• International Journal of Internet, Broadcasting and Communication
• /
• v.12 no.2
• /
• pp.45-50
• /
• 2020

A Language model with neural networks commonly trained with likelihood loss. Such that the model can learn the sequence of human text. State-of-the-art results achieved in various language generation tasks, e.g., text summarization, dialogue response generation, and text generation, by utilizing the language model's next token output probabilities. Monotonous and boring outputs are a well-known problem of this model, yet only a few solutions proposed to address this problem. Several decoding techniques proposed to suppress repetitive tokens. Unlikelihood training approached this problem by penalizing candidate tokens probabilities if the tokens already seen in previous steps. While the method successfully showed a less repetitive generated token, the method has a large memory consumption because of the training need a big vocabulary size. We effectively reduced memory footprint by encoding words as sequences of subword units. Finally, we report competitive results with token level unlikelihood training in several automatic evaluations compared to the previous work.

• Smart Media Journal
• /
• v.7 no.4
• /
• pp.17-23
• /
• 2018

We proposed a LSTM-based RNN model that can effectively perform the automatic spacing characteristics. For those long or noisy sentences which are known to be difficult to handle within Neural Network Learning, we defined a proper input data format and decoding data format, and added dropout, bidirectional multi-layer LSTM, layer normalization, and attention mechanism to improve the performance. Despite of the fact that Sejong corpus contains some spacing errors, a noise-robust learning model developed in this study with no overfitting through a dropout method helped training and returned meaningful results of Korean word spacing and its patterns. The experimental results showed that the performance of LSTM sequence-to-sequence model is 0.94 in F1-measure, which is better than the rule-based deep-learning method of GRU-CRF.

• KSII Transactions on Internet and Information Systems (TIIS)
• /
• v.16 no.1
• /
• pp.60-79
• /
• 2022

The U-Net architecture-based segmentation models attained remarkable performance in numerous medical image segmentation missions like skin lesion segmentation. Nevertheless, the resolution gradually decreases and the loss of spatial information increases with deeper network. The fusion of adjacent layers is not enough to make up for the lost spatial information, thus resulting in errors of segmentation boundary so as to decline the accuracy of segmentation. To tackle the issue, we propose a new deep learning-based segmentation model. In the decoding stage, the feature channels of each decoding unit are concatenated with all the feature channels of the upper coding unit. Which is done in order to ensure the segmentation effect by integrating spatial and semantic information, and promotes the robustness and generalization of our model by combining the atrous spatial pyramid pooling (ASPP) module and channel attention module (CAM). Extensive experiments on ISIC2016 and ISIC2017 common datasets proved that our model implements well and outperforms compared segmentation models for skin lesion segmentation.

• Annual Conference on Human and Language Technology
• /
• 2015.10a
• /
• pp.115-118
• /
• 2015

Neural Machine Translation (NMT) 모델은 단일 신경망 구조만을 사용하는 End-to-end 방식의 기계번역 모델로, 기존의 Statistical Machine Translation (SMT) 모델에 비해서 높은 성능을 보이고, Feature Engineering이 필요 없으며, 번역 모델 및 언어 모델의 역할을 단일 신경망에서 수행하여 디코더의 구조가 간단하다는 장점이 있다. 그러나 NMT 모델은 출력 언어 사전(Target Vocabulary)의 크기에 비례해서 학습 및 디코딩의 속도가 느려지기 때문에 출력 언어 사전의 크기에 제한을 갖는다는 단점이 있다. 본 논문에서는 NMT 모델의 출력 언어 사전의 크기 제한 문제를 해결하기 위해서, 입력 언어는 단어 단위로 읽고(Encoding) 출력 언어를 문자(Character) 단위로 생성(Decoding)하는 방법을 제안한다. 출력 언어를 문자 단위로 생성하게 되면 NMT 모델의 출력 언어 사전에 모든 문자를 포함할 수 있게 되어 출력 언어의 Out-of-vocabulary(OOV) 문제가 사라지고 출력 언어의 사전 크기가 줄어들어 학습 및 디코딩 속도가 빨라지게 된다. 실험 결과, 본 논문에서 제안한 방법이 영어-일본어 및 한국어-일본어 기계번역에서 기존의 단어 단위의 NMT 모델보다 우수한 성능을 보였다.

• Journal of Broadcast Engineering
• /
• v.26 no.7
• /
• pp.868-876
• /
• 2021

Deploying deep neural network model training performs remarkable performance in the fields of Object detection and Instance segmentation. To train these models, features are first extracted from the input image using a backbone network. The extracted features can be reused by various tasks. Research has been actively conducted to serve various tasks by using these learned features. In this process, standardization discussions about encoding, decoding, and transmission methods are proceeding actively. In this scenario, it is necessary to analyze the response characteristics of features against various distortions that may occur in the data transmission or data compression process. In this paper, experiment was conducted to inject various distortions into the feature in the object recognition task. And analyze the mAP (mean Average Precision) metric between the predicted value output from the neural network and the target value as the intensity of various distortions was increased. Experiments have shown that features are more robust to distortion than images. And this points out that using the feature as transmission means can prevent the loss of information against the various distortions during data transmission and compression process.

• Journal of information and communication convergence engineering
• /
• v.1 no.4
• /
• pp.177-182
• /
• 2003

Conventional multilayer feedforward artificial neural networks are very effective in dealing with spatial problems. To deal with problems with time dependency, some kinds of memory have to be built in the processing algorithm. In this paper we show how the newly proposed Serial Input Neuron (SIN) convolutional decoders can be derived. As an example, we derive the SIN decoder for rate code with constraint length 3. The SIN is tested in Gaussian channel and the results are compared to the results of the optimal Viterbi decoder. A SIN approach to decode convolutional codes is presented. No supervision is required. The decoder lends itself to pleasing implementations in hardware and processing codes with high speed in a time. However, the speed of the current circuits may set limits to the codes used. With increasing speeds of the circuits in the future, the proposed technique may become a tempting choice for decoding convolutional coding with long constraint lengths.

• The Journal of Korean Institute of Communications and Information Sciences
• /
• v.21 no.8
• /
• pp.2109-2118
• /
• 1996

The neural convolutional decoder(NCD) was proposed as a method of decoding convolutional codes. In this paper, simulation results are presented for coherent BPSK in memoryless AWGN channels and coherent QPSK in the satellite channels. The NCD can learn the nonlinear distortion caused by the charactersitics of the satellite channel including the filtering effects and the nonlinear effects of the travling wave tube amplifier(TWTA). Thus, as compared with the AWGN channel, the performance difference in the satellite channel between the NCD for the systematic code and the Viterbi decoder for the nonsystematic code is reduced.