International Journal of Smart Computing and Artificial Intelligence

Rotation Weight Update:

Tetsuya Hori, Yuki Sekiya, Yoichi Takenaka — Sat, 30 Nov 2024 11:57:20 +0000

Numerous approaches have been developed to enhance the capabilities of deep learning in image recognition. We propose a method called ”Rotational-update,” which cyclically updates the weight of neurons. This approach segments the neurons in a fully connected layer into groups of equal size, each containing √N neurons, with N representing the total neuron count in the layer. It selectively updates the weights of one group at a time per mini-batch. This selective updating mechanism aims to curb excessive learning, potentially reducing overfitting and enhancing validation accuracy. A notable aspect of this method is its compatibility for concurrent use with other techniques, including batch-normalization and dropout.
We used the CIFAR10 dataset for image recognition tasks to validate the method’s efficacy, employing three neural network architectures: VGG-16, ResNet-110, and ResNet-152. Our findings indicate that integrating our proposed method with batch normalization outperforms the accuracy of the combination of dropout and batch normalization. Specifically, the proposed Rotational-update method achieved an accuracy improvement of up to 5 percentage points in VGG-16 and one percentage point in ResNet-110 compared to traditional methods. Thus, we deduce that substituting dropout with our proposed method enhances image recognition task performance and reduces overfitting.

Classroom Utterance Analysis and Visualization Using a Generative Deep Neural Networks for Dialogue Model

Sakuei Onishi, Tomohiko Yasumori, Hiromitsu Shiina — Tue, 29 Oct 2024 13:40:14 +0000

In elementary school and other classes at different levels, teachers have less time for reflection, which can involve self-assessment and classroom observation. Never-theless, reflection activities are becoming increasingly important. Unfortunately, few studies support teachers ’reflection, and studies that analyze classroom utterances treat them as text and do not consider them as dialogues. However, in the field of dialogue response generation, some dialogue models using neural networks have been proposed. In this study, we propose a dialogue model that considers the domain of the class and extracts similar utterances. The proposed model considers the domain of ele-mentary school classes, a domain in which speakers can be classified, and incorporates a method to abstract the characteristics of speakers by clustering. The proposed model can be constructed with a relatively small number of parameters. We also developed a system to visualize the classification probabilities analyzed using the proposed dia-logue model. The developed visualization system was evaluated by experts and was found to visually recognize the classification bias of an utterance and to confirm the quality of the utterance by the similarity of the utterances.