• 방송공학회논문지
• /
• 제25권5호
• /
• pp.742-749
• /
• 2020

본 논문에서는 하나의 공용 신경망을 사용하여 음소와 감정을 모두 인식하는 방법과 공용 신경망 학습을 위한 다중 학습 방법을 제안한다. 공용 신경망은 동일한 동작을 수행하여 두 정보를 모두 인식하며, 이는 인간이 하나의 청각기관으로 여러 정보를 동시에 인식하는 구조에 해당한다. 다중 학습은 여러 정보를 위한 공통 모델링을 진행하므로 여러 정보에 대한 일반화된 학습을 진행시켜 기존의 정보별 개별 학습에서 나타나는 과적합을 감소시키고 인식 성능을 향상시킨다. 또한, 다중 학습에서 음소 인식에 가중치를 부여하여 음소 인식 성능을 추가 향상시키는 방법을 제안한다. 동일한 특성벡터와 신경망을 사용할 때, 제안한 다중 학습이 적용된 공용 신경망의 성능이 각 정보별로 학습시킨 개별 신경망에 비하여 우수한 것을 확인하였다.

• Journal of Information Processing Systems
• /
• 제17권3호
• /
• pp.441-452
• /
• 2021

Federated learning provides an efficient integrated model for distributed data, allowing the local training of different data. Meanwhile, the goal of multi-task learning is to simultaneously establish models for multiple related tasks, and to obtain the underlying main structure. However, traditional federated multi-task learning models not only have strict requirements for the data distribution, but also demand large amounts of calculation and have slow convergence, which hindered their promotion in many fields. In our work, we apply the rank constraint on weight vectors of the multi-task learning model to adaptively adjust the task's similarity learning, according to the distribution of federal node data. The proposed model has a general framework for solving optimal solutions, which can be used to deal with various data types. Experiments show that our model has achieved the best results in different dataset. Notably, our model can still obtain stable results in datasets with large distribution differences. In addition, compared with traditional federated multi-task learning models, our algorithm is able to converge on a local optimal solution within limited training iterations.

• Journal of Multimedia Information System
• /
• 제9권2호
• /
• pp.75-86
• /
• 2022

The existing image quality assessment (IQA) datasets have a small number of samples. Some methods based on transfer learning or data augmentation cannot make good use of image quality-related features. A No Reference (NR)-IQA method based on multi-task training and quality awareness is proposed. First, single or multiple distortion types and levels are imposed on the original image, and different strategies are used to augment different types of distortion datasets. With the idea of weak supervision, we use the Full Reference (FR)-IQA methods to obtain the pseudo-score label of the generated image. Then, we combine the classification information of the distortion type, level, and the information of the image quality score. The ResNet50 network is trained in the pre-train stage on the augmented dataset to obtain more quality-aware pre-training weights. Finally, the fine-tuning stage training is performed on the target IQA dataset using the quality-aware weights to predicate the final prediction score. Various experiments designed on the synthetic distortions and authentic distortions datasets (LIVE, CSIQ, TID2013, LIVEC, KonIQ-10K) prove that the proposed method can utilize the image quality-related features better than the method using only single-task training. The extracted quality-aware features improve the accuracy of the model.

• 한국정보처리학회:학술대회논문집
• /
• 한국정보처리학회 2022년도 추계학술발표대회
• /
• pp.308-310
• /
• 2022

With the rise of the Al wave, task-oriented dialogue systems have become one of the popular research directions in academia and industry. Currently, task-oriented dialogue systems mainly adopt pipelined form, which mainly includes natural language understanding, dialogue state decision making, dialogue state tracking and natural language generation. However, pipelining is prone to error propagation, so many task-oriented dialogue systems in the market are only for single-round dialogues. Usually single- domain dialogues have relatively accurate semantic understanding, while they tend to perform poorly on multi-domain, multi-round dialogue datasets. To solve these issues, we developed a paddle-based multi-domain task-oriented Chinese dialogue system. It is based on NEZHA-base pre-training model and CrossWOZ dataset, and uses intention recognition module, dichotomous slot recognition module and NER recognition module to do DST and generate replies based on rules. Experiments show that the dialogue system not only makes good use of the context, but also effectively addresses long-term dependencies. In our approach, the DST of dialogue tracking state is improved, and our DST can identify multiple slotted key-value pairs involved in the discourse, which eliminates the need for manual tagging and thus greatly saves manpower.

• 영어어문교육
• /
• 제16권1호
• /
• pp.123-142
• /
• 2009

The present study used multi-faceted Rasch measurement to explore the characteristics and bias patterns of non-native raters when they scored L2 writing tasks. Three raters scored 254 writing tasks written by Korean university students on two topics adapted from the TOEFL Test of Written English (TWE). The written products were assessed using a five-category rating scale (Content, Organization, Language in Use, Grammar, and Mechanics). The raters only showed a difference in severity with regard to rating categories but not in task types. Overall, the raters scored Grammar most harshly and Organization most leniently. The results also indicated several bias patterns of ratings with regard to the rating categories and task types. In rater-task bias interactions, each rater showed recurring bias patterns in their rating between two writing tasks. Analysis of rater-category bias interaction showed that the three raters revealed biased patterns across all the rating categories though they were relatively consistent in their rating. The study has implications for the importance of rater training and task selection in L2 writing assessment.

• 한국음향학회지
• /
• 제40권5호
• /
• pp.515-522
• /
• 2021

음성감정인식을 위한 훈련 데이터는 감정 레이블링의 어려움으로 인해 충분히 확보하기 어렵다. 본 논문에서는 음성감정인식의 성능 개선을 위해 트랜스포머 기반 모델에 대규모 음성인식용 훈련 데이터를 통한 전이학습을 적용한다. 또한 음성인식과의 다중작업학습을 통해 별도의 디코딩 없이 문맥 정보를 활용하는 방법을 제안한다. IEMOCAP 데이터 셋을 이용한 음성감정인식 실험을 통해, 가중정확도 70.6 % 및 비가중정확도 71.6 %를 달성하여, 제안된 방법이 음성감정인식 성능 향상에 효과가 있음을 보여준다.

• International Journal of Internet, Broadcasting and Communication
• /
• 제16권3호
• /
• pp.192-198
• /
• 2024

This paper investigates the application of multi-agent deep reinforcement learning in the fighting game Samurai Shodown using Proximal Policy Optimization (PPO) and Advantage Actor-Critic (A2C) algorithms. Initially, agents are trained separately for 200,000 timesteps using Convolutional Neural Network (CNN) and Multi-Layer Perceptron (MLP) with LSTM networks. PPO demonstrates superior performance early on with stable policy updates, while A2C shows better adaptation and higher rewards over extended training periods, culminating in A2C outperforming PPO after 1,000,000 timesteps. These findings highlight PPO's effectiveness for short-term training and A2C's advantages in long-term learning scenarios, emphasizing the importance of algorithm selection based on training duration and task complexity. The code can be found in this link https://github.com/Lexer04/Samurai-Shodown-with-Reinforcement-Learning-PPO.

• 한국멀티미디어학회논문지
• /
• 제22권10호
• /
• pp.1149-1159
• /
• 2019

We present a novel deep learning architecture for obtaining a latent image from a single blurry image, which contains dynamic motion blurs through object/camera movements. The proposed architecture consists of two sub-modules: blur image restoration and optical flow estimation. The tasks are highly related in that object/camera movements make cause blurry artifacts, whereas they are estimated through optical flow. The ablation study demonstrates that training multi-task architecture simultaneously improves both tasks compared to handling them separately. Objective and subjective evaluations show that our method outperforms the state-of-the-arts deep learning based techniques.

• 디지털융복합연구
• /
• 제17권12호
• /
• pp.243-248
• /
• 2019

다중작업학습(Multi-Task Learning, MTL) 기법은 하나의 신경망을 통해 다양한 작업을 동시에 수행하고 각 작업 간에 상호적으로 영향을 미치면서 학습하는 방식을 말한다. 본 연구에서는 전통문화 말뭉치를 직접 구축 및 학습데이터로 활용하여 다중작업학습 기법을 적용한 개체명 인식 모델에 대해 성능 비교 분석을 진행한다. 학습 과정에서 각각의 품사 태깅(Part-of-Speech tagging, POS-tagging) 과 개체명 인식(Named Entity Recognition, NER) 학습 파라미터에 대해 Bi-LSTM 계층을 통과시킨 후 각각의 Bi-LSTM을 계층을 통해 최종적으로 두 loss의 joint loss를 구한다. 결과적으로, Bi-LSTM 모델을 활용하여 단일 Bi-LSTM 모델보다 MTL 기법을 적용한 모델에서 1.1%~4.6%의 성능 향상이 있음을 보인다.

• 음성과학
• /
• 제11권1호
• /
• pp.7-20
• /
• 2004

In this paper, we propose a two-step noise compensation algorithm in feature extraction for achieving robust speech recognition. The proposed method frees us from requiring a priori information on noisy environments and is simple to implement. First, in frequency domain, the Harmonics-based Spectral Subtraction (HSS) is applied so that it reduces the additive background noise and makes the shape of harmonics in speech spectrum more pronounced. We then apply a judiciously weighted variance Feature Vector Normalization (FVN) to compensate for both the channel distortion and additive noise. The weighted variance FVN compensates for the variance mismatch in both the speech and the non-speech regions respectively. Representative performance evaluation using Aurora 2 database shows that the proposed method yields 27.18% relative improvement in accuracy under a multi-noise training task and 57.94% relative improvement under a clean training task.