榎原 学人 | M, 2回目発表 | 数理情報学(計算神経科学) | 池田 和司☆ | 川鍋 一晃 | 杉本 徳和 | 田中 沙織 |
title: Covert speech decoding using optical pumping magnetometer (OPM)
abstract: Amyotrophic lateral sclerosis (ALS) is a neurodegenerative disease characterized by muscle weakness resulting from the progressive damage to motor neurons responsible for muscle control. This study aims to develop a non-invasive brain-machine interface (BMI) for ALS patients who retain cognitive function but struggle with verbal communication. The proposed BMI utilizes an optical pumping magnetometer (OPM) to decode covert speech—words that patients are able to think but not vocalize—thereby offering a risk-free communication solution. language of the presentation: Japanese | ||||||
吉田 雄丸 | M, 2回目発表 | 数理情報学(計算神経科学) | 池田 和司☆ | 川鍋 一晃 | 杉本 徳和 | 田中 沙織 |
title: Finer EEG-based Emotion Recognition with Semantic Space Theory
abstract: This research plans to research EEG-based emotion recognition based on the Semantic Space Theory of emotions. This research focuses on how finely distinguishing emotions impacts stress resilience, and is developing novel methods to classify emotions more precisely using EEG data. In this experiment, brainwaves are measured while participants watch video stimuli, using this data to explore the relationship between emotional diversity and brain activity patterns. This approach has the potential to offer more detailed insights into emotion recognition compared to traditional methods. Furthermore, Since EEG and emotion rating are multidimensional data, I aim to demonstrate whether there is a correlation between them through multivariate analysis and to gain a new understanding of emotion recognition. language of the presentation: Japanese | ||||||
LI SHANGLIN | M, 2回目発表 | 数理情報学(計算神経科学) | 池田 和司☆ | 川鍋 一晃 | 杉本 徳和 | 田中 沙織 |
山口 晴久 | M, 1回目発表 | 数理情報学(計算神経科学) | 池田 和司☆ | 作村 諭一 | 川鍋 一晃 | 田中 沙織 |
中畔 彪雅 | M, 2回目発表 | 自然言語処理学(ロボット対話知能) | 渡辺 太郎☆ | 吉野 幸一郎 | 河野 誠也 |
title: Constructing Speech-to-Text dialogue model using continuous features
abstract: Many existing spoken dialogue systems are realized by connecting speech recognition and dialogue system modules in a cascade. However, such cascade models requires a process of converting speech to text, which lost speech information is necessary for natural spoken dialogue. Therefore, we aim to construct the Speech-to-Text dialogue model involving continuous features that contain richer information than text. language of the presentation: Japanese 発表題目: 連続表現を用いたSpeech-to-Text対話モデルの構築 概要:既存の音声対話システムの多くは音声認識,対話生成のモジュールをカスケード型に接続することで実現される.しかし,このようなカスケード型モデルでは,音声をテキストに変換する過程が必要であり,自然な音声対話の実現に必要なテキスト化できない音声情報は欠落する問題がある.そこで,本研究ではテキスト以上に豊かな情報を持つ連続表現を用いた音声—テキスト対話モデルの構築を目指した. | |||||
日浦 隆博 | M, 2回目発表 | 自然言語処理学(ロボット対話知能) | 渡辺 太郎☆ | 吉野 幸一郎 | 河野 誠也 |
title: Symbolic Knowledge Distillation Using Adversarial Learning
abstract: An approach that combines knowledge graphs and large language models (LLMs) is gaining attention to improve the accuracy of knowledge inference in LLMs. In particular, symbolic knowledge distillation transfers LLM outputs to a knowledge graph and uses them for training knowledge inference models. However, since the outputs may contain errors, filtering is necessary. This study proposes a framework where a filter model and an evaluation model are trained adversarially to filter LLM outputs using only existing knowledge graphs, thereby expanding the knowledge graph. language of the presentation: Japanese 発表題目: 敵対的学習を用いた記号的知識蒸留 発表概要: 大規模言語モデル(LLM)の知識推論における精度向上を目指し、知識グラフとLLMを組み合わせたアプローチが注目されている。特に記号的知識蒸留はLLMの出力を知識グラフに転移し、知識推論モデルの学習に利用するものであるが、出力には誤りが含まれるためフィルタが必要である。本研究ではフィルタモデルと評価モデルを敵対的に学習させることで、既存の知識グラフのみからLLMの出力をフィルタし、知識グラフを大規模化する枠組みを提案する。 | |||||
藤田 一天 | M, 2回目発表 | 自然言語処理学(ロボット対話知能) | 渡辺 太郎☆ | 吉野 幸一郎 | 河野 誠也 |
title: Recognition of Plan Fulfillment Based on Granularity Alignment in Action Recognition
abstract: Through situational awareness using techniques such as captioning, it has become possible to describe user actions in video sequences using natural language. By employing such frameworks, it is feasible to develop applications for tasks such as determining the fulfillment of planned actions. However, when trying to assess the fulfillment of plans based on recognized user actions expressed in language, the granularity between the recognized actions described in text and the planned actions often differs, and there may also be ambiguous expressions. Therefore, in this study, we use a paraphrasing model to break down and align each text to the same level of granularity. By aligning these subdivided actions, we aim to construct a model that accurately recognizes the fulfillment of plans. language of the presentation: Japanese 発表題目: 行動認識の粒度アライメントに基づく予定の履行認識 発表概要: キャプショニングなどを用いた状況認識により、動画像中のユーザ行動を自然言語で表現することが可能になった。こうした枠組みを用いて、行動予定の履行判断などのアプリケーションを実現することが可能である。しかし、言語で表現されたユーザ行動の認識結果を用いて予定の履行を判断しようとした時、テキストで表現される行動の認識結果と行動予定はしばしば粒度が異なり、また曖昧な表現も存在する。そこで本研究では言い換えモデルを用いてそれぞれのテキストを細分化して粒度に揃える。この細分化された行動間のアライメントを取ることで、予定の履行を適切に認識するモデルの構築を行う。 | |||||
三輪 拓真 | M, 2回目発表 | 自然言語処理学(ロボット対話知能) | 渡辺 太郎☆ | 吉野 幸一郎 | 河野 誠也 |
title: Investigation about direct model using quantum computing
abstract: The next stage of deep learning is expected to be applied to complex tasks with social implementation in mind. There are two approaches to such tasks: the Cascade Model and the End-to-End Model. While the cascade model is highly versatile, information loss occurs at intermediate outputs. On the other hand, the End-to-End model has no information loss, but its use is limited and training data is not abundant. To solve these problems, we focused on the property of quantum computation that allows each model to be learned independently but can be combined during inference. We propose a direct model that uses quantum computation to reduce the learning cost of the End-to-End model while still performing inference directly. language of the presentation: Japanese | |||||