|
主持人(PI)
王新民
Hsin-Min Wang
|
研究主題(Research Topic)
語音處理
Speech Processing
|
研究介紹(Introduction)
我們致力於符合我國語言使用語境（國語、臺語、客語、原住民語、英語）的語音處理研究，包括語音辨識、語音合成/轉換、語音翻譯、大語言模型及各種應用。另外，針對各種言語障礙，例如電子喉語音和構音障礙語音，我們希望利用語音處理技術來提升語音品質及可懂度。我們的研究兼重學術發表和系統開發。
We are dedicated to speech processing research that aligns with our country's language usage context (Mandarin, Taiwanese, Hakka, Indigenous languages, and English), including speech recognition, speech synthesis/conversion, speech translation, large language models, and various applications. Furthermore, we aim to utilize speech processing technologies to improve speech quality and intelligibility for various speech disorders, such as electrolaryngeal speech and dysarthric speech. Our research emphasizes both academic publication and system development.
|
其他資訊(Other Information)
|
|
主持人(PI)
廖純中
Churn-Jung Liau
|
研究主題(Research Topic)
應用邏輯
Applied Logic
|
研究介紹(Introduction)
我們對探討符號邏輯(特別是非古典邏輯)理論及其在各領域的應用上都有興趣。
We are widely interested in the theory of symbolic logic (especially non-classical logics) and its applications to various domains.
|
其他資訊(Other Information)
|
|
主持人(PI)
洪鼎詠
Ding-Yong Hong
|
研究主題(Research Topic)
深度學習軟體與硬體協同優化研究
Deep Learning Software/Hardware Co-optimization
|
研究介紹(Introduction)
我們將研究深度學習軟體與硬體協同優化方法。(1) 研究如何利用編譯器技術, 優化深度學習模型, 使其在CPU/GPU/AI加速器上達到最佳的運算效能。(2) 針對壓縮模型(pruning/quantization), 設計深度學習模型architecture/compiler/parallelization優化方案。
We aim to study hardware/software co-optimization for deep learning models. (1) Exploiting compiler techniques to accelerate deep learning applications on CPUs/GPUs/AI accelerators. (2) Enhancing compressed models (pruning/quantization) with compiler and parallelization techniques.
|
其他資訊(Other Information)
|
|
主持人(PI)
蕭邱漢
Chiu-Han Hsiao
|
研究主題(Research Topic)
基於綠色學習與聯邦學習架構，應用於整合醫學影像和臨床特徵識別良性和惡性腎細胞癌
Green Learning–Driven Federated Framework for Integrating Medical Imaging and Clinical Features in Benign and Malignant RCC Identification
|
研究介紹(Introduction)
準確區分良性和惡性腎細胞癌 (RCC) 仍然是一項重要的臨床挑戰。儘管 CT 影像等非侵入性放射學技術可以表示腫瘤特徵，但辦讀上仍依賴臨床醫師耗時費力的手動分析。本計畫將醫學影像和放射組學資料整合到一個醫療保健系統的智慧 AI 協調器中，強調監督學習（即監督的力量）在指導領域資訊模型研發。本系統融合綠色學習 (GL) 方法和數學優化法，以提高計算效率和可持續性。此外，另設計聯邦學習 (FL) 架構，用以支援保護隱私的多機構協作，從而為臨床醫師提供自適應且可解釋的決策支援工作流程。目標在提高診斷精度，簡化臨床評估，並為 RCC 的評估和治療規劃提供輔助參考。
Accurately distinguishing between benign and malignant renal cell carcinoma (RCC) remains a critical clinical challenge. Although noninvasive radiological modalities such as computed tomography (CT) can reveal tumor characteristics, image interpretation continues to rely heavily on time-consuming and labor-intensive manual analysis by clinicians. This project integrates medical imaging and radiomics data into an intelligent AI orchestrator within a healthcare system, emphasizing the role of supervised learning, the "power of supervision" in guiding domain-informed model development.
The proposed framework incorporates Green Learning (GL) strategies and mathematical optimization techniques to enhance computational efficiency and sustainability. In addition, a Federated Learning (FL) architecture is designed to support privacy-preserving, multi-institutional collaboration, enabling secure model training without centralized data sharing. Together, these components provide clinicians with an adaptive and interpretable decision-support workflow.
The overall objective is to improve diagnostic accuracy, streamline clinical evaluation, and offer reliable decision support for RCC assessment and treatment planning.
|
其他資訊(Other Information)
|
|
主持人(PI)
陳孟彰
Meng Chang Chen
|
研究主題(Research Topic)
應用大型語言模型與深度學習於 APT 攻擊活動偵測
APT Campaign Detection via LLM and Deep Learning Techniques
|
研究介紹(Introduction)
本計畫擬研究大型語言模型（LLM）與深度學習模式技術，以提升進階持續性威脅（APT）攻擊活動的偵測能力。透過建模系統稽核紀錄、事件序列與技術層級抽象（如 MITRE ATT&CK），本研究可有效辨識、分解並歸因於具高度隱匿性且多階段的 APT 攻擊行為，克服傳統規則式與特徵碼式防禦方法的限制。
在此架構中，LLM 負責提供語意推理、情境理解與跨事件關聯分析，而深度學習模型則用以捕捉長時間尺度下的行為模式與異常特徵。所提出之方法可實現具韌性的攻擊活動層級偵測，支援對未知或持續演化之攻擊技術的識別，並提升自動化威脅獵捕系統在真實複雜環境中的可解釋性與適應能力。
This project aims to advance the detection of Advanced Persistent Threat (APT) campaigns by integrating Large Language Models (LLMs) with deep learning–based behavioral analysis. By modeling system audit logs, event sequences, and technique-level abstractions (e.g., MITRE ATT&CK), the project seeks to identify, disentangle, and attribute stealthy multi-stage APT activities that evade traditional rule-based and signature-driven defenses. LLMs are leveraged to provide semantic reasoning, contextual understanding, and cross-event correlation, while deep learning models capture temporal patterns and anomalous behaviors at scale. The proposed framework enables robust campaign-level detection, supports unseen or evolving attack techniques, and improves the interpretability and adaptability of automated threat-hunting systems in complex, real-world environments.
|
其他資訊(Other Information)
|
|
主持人(PI)
陳伶志
Ling-Jyh Chen
|
研究主題(Research Topic)
針對網路管理及特定領域知識庫之 SLM 微調與跨平台應用研究
Fine-tuning and Deploying SLMs for Network Administration and Specialized Knowledge Bases
|
研究介紹(Introduction)
在生成式 AI 的應用藍圖中，通用型 LLM 往往難以直接應對高度專業化的技術領域。今年，我們的暑期實習專案將聚焦於 SLM (Small Language Models) 的垂直領域開發，目標是打造出能精準理解專業知識、具備高度隱私性且能跨平台運作的 AI 助理。
本計畫將以 「網路管理（Network Administration）」 為核心應用情境——包含處理複雜的網路協議、自動化維運腳本（Scripts）與故障排除——同時也歡迎對其他特定領域專業知識庫（如法律、醫療、或企業內部私有知識庫）有興趣的同學加入。我們強調模型的「場域落地價值」，實習生將挑戰如何透過微調技術（Fine-tuning），使 SLM 能在雲端伺服器或行動裝置（如維運工程師的手持設備）上高效運行。
我們的研究內容兼具技術挑戰與實作意義，包含：
- 垂直領域微調： 針對網路維運資料或特定領域技術文件進行 SFT、DPO 等微調，建立專家級的小型化模型。
- 專業知識庫整合： 結合檢索增強生成（RAG）技術，解決模型在特定領域的「幻覺」問題，確保回覆的精準度與權威性。
- 輕量化部署研究： 探索模型量化（Quantization）與蒸餾技術，達成 SLM 在行動端與邊緣設備的低延遲推論。
- 真實場域驗證： 開發可操作的原型系統，驗證 SLM 在專業辦公環境或自動化維運流程中的實際表現。
我們期待您具備對 AI 領域的熱情、熟悉模型微調工具鏈（如 Hugging Face, PyTorch）、具有出色的邏輯分析能力。如果您渴望將最前沿的語言模型技術轉化為解決特定領域問題的實戰工具，歡迎加入我們。
In the landscape of Generative AI, general-purpose LLMs often struggle with highly specialized technical domains. This year, our internship program focuses on the vertical development of Small Language Models (SLMs), aiming to create AI assistants that precisely understand professional knowledge, ensure data privacy, and operate across diverse platforms.
The primary application scenario for this project will be Network Administration (encompassing complex protocols, NetOps automation, and troubleshooting), but it also extends to other specialized knowledge bases (such as legal, medical, or private corporate domains). We emphasize "real-world operability," where interns will explore fine-tuning techniques to enable SLMs to run efficiently on Cloud Servers or Mobile Devices for on-site professional use.
Our research encompasses technical innovation and practical implementation:
- Vertical Domain Fine-tuning: Performing SFT and DPO on network data or technical documentation to build expert-level lightweight models.
- Knowledge Base Integration: Utilizing RAG techniques to mitigate hallucinations and ensure the accuracy of domain-specific responses.
- Lightweight Deployment: Exploring quantization and distillation to achieve low-latency inference on mobile and edge devices.
- Real-World Validation: Developing prototypes to verify SLM performance in professional office environments or automated maintenance workflows.
We seek candidates with a passion for AI, hands-on experience with fine-tuning frameworks (e.g., Hugging Face, PyTorch), and strong analytical skills. If you are eager to transform cutting-edge language models into practical tools for specialized domains, we welcome you to join our team.
|
其他資訊(Other Information)
|
|
主持人(PI)
林仲彥
Chung-Yen Lin
|
研究主題(Research Topic)
自然語言人工智慧生醫資料對話解析
Conversational AI for Biomedical Big Data Analytics
|
研究介紹(Introduction)
我們的團隊主要研究模式與非模式生物之多維基因體學（OMICS），並利用生物序列語言模型與大語言模型等，來與包括基因體、轉錄體、單細胞轉錄體、蛋白質交互網路、腸道微生物與疾病關連等巨量資訊數據進行對話與解析。目前致力利用人工智慧模型，以台灣人體資料庫為基礎，結合先天遺傳差異、身體檢測數值與後天環境等，來以全新的視角，來建立預測模型與對話平台，希望能早期預防及解析老化與疾病等相關問題。我們的成員來自資料科學、生物醫學與資訊技術等各類專業領域，是一個跨領域的研究團隊，歡迎不同背景（資訊、統計、數學及生物相關）的人才一起合作。本團隊研究範圍以基因體組裝解析、水生經濟動物基因體育種、精準健康老化、病原智慧分型、新型抗菌/抗病毒藥物的開發篩選與合成驗證、及利用人類腸道與環境微生物來進行人工智慧疾病與治療成效預測等課題為主，同時發展新的高速計算工具及雲端分析平台，以及引入深度學習等策略，來探討基因、病原與環境的三角互動關係。
Our team focuses on multidimensional genomics (OMICS) research, combining biological sequence language models, large language models (LLMs), and AI to analyze large datasets across genomics, transcriptomics, microbiota, and disease associations. Using the Taiwan Biobank, we integrate genetic, clinical, and environmental data to develop predictive models and dialogue platforms for early disease prevention and aging research.
We are a multidisciplinary team that welcomes experts from diverse fields, including data science, biology, and IT. Key research areas include single-cell analysis in full-length transcripts, aquatic animal breeding, precision aging, pathogen typing, drug development, and AI-based disease prediction using microbiota. We also create advanced computational tools and cloud platforms to explore gene-pathogen-environment interactions.
|
其他資訊(Other Information)
|
|
主持人(PI)
楊得年
De-Nian Yang
|
研究主題(Research Topic)
代理式AI之多模態資料探勘與多媒體延展實境網路優化
Agentic AI for Multimodal Data Mining and XR Multimedia Network Optimization
|
研究介紹(Introduction)
(一)面向AI代理時代的資料探勘、機器學習、基礎模型與演算法設計：
1. 基於虛擬、擴增、延展實境(VR/AR/XR)的推薦系統：如規劃避免3D暈眩或撞到障礙物的虛實路徑、基於多模態基礎模型之畫面內容推薦以最大化社群共感和個人喜好、結合知識增強（RAG）之基於LLM與情境之主動式推薦、具代理式決策之主動互動推薦、具因果推論之偏好學習與推薦、具AI倫理之NFT交易推薦系統、及以合成資料驅動之XR情境推薦測試與訓練資料建置。
2. 社群影響力分析與優化：如以AI基礎模型為核心之多面向社群影響力學習與預測、基於生成式AI之動態社群網路生成模型、結合多模態基礎模型之虛實世界社群分析、具因果推論之影響力學習、具AI倫理之影響力最大化、個人化密度彈性群體查詢、基於圖推理之子結構資訊融合、具可解釋性之GNN反事實學習、及具代理式探索之社群擴散最佳化。
3. 其他應用領域的推薦系統：如結合知識增強（RAG）之LLM推薦、具代理式工作流之多步驟個人化推薦、具可解釋性之推薦與探勘、具因果推論之推薦成效歸因、推薦系統多人毒害攻擊、具可解釋性之異質性推薦系統異常偵測、以混合專家模型支援之大規模推薦推論與服務、以合成資料驅動之攻擊/異常情境生成與修正、群組探勘優化、與結合AI代理之活動潛在參與者推薦。
(二)次世代網路演算法設計與分析：
藉由分析問題NP困難度及不可近似性的方法，以及高階演算法設計技巧 (如近似演算法、競爭演算法、AI演算法等)，來解決多媒體網路中的各類應用問題。
1. 延展實境網路：如規劃有線及無線網路資源配置和排程方式、選定3D多視角影片傳輸及合成之場景、決定3D合成相關參數和虛擬實境頭盔使用者暈眩減緩機制之設計、並結合語意通訊，以最佳化多媒體網路傳輸效率及確保使用者的沉浸體驗。
2. 低軌衛星網路：結合群播流量工程，6G和衛星裝置間的直接通訊，包含無人機、衛星和地面綜合網路，並考慮網路中的能源效率和永續性。
3. 行動邊緣運算網路：如結合生成式數位雙生 (Generative Digital Twin)，設計高階演算法以建置高效、可靠的社群物聯網和群眾外包系統，並利用生成式AI補全缺漏的感測資料，並在邊緣端模擬「虛擬情境」進行演算法測試，減少對真實資料的依賴。
4. AI網路中的各類優化問題：如在不同AI訓練框架下 (例如: 聯盟式學習和圖神經網路)，設計動態路由、選擇資料源、選擇訓練特徵及拓樸控制，以最小化總頻寬和計算資源消耗，並確保線路／節點容量限制及不同應用需求。
延展實境是整合多個虛擬世界的系統，讓人們透過虛擬化身在裡面社交、購物和創作。現實世界的物品和服務也以數位雙生的方式存在，成為實體裝置和服務的虛擬代表，連接真實世界和虛擬世界。在此基礎上，AI正以基礎模型（Foundation Models）為核心重塑延展實境中的內容生產與互動方式（例如生成式內容、多模態理解、個人化體驗與代理式AI（Agentic AI）驅動的主動互動）。因此，下一代延展實境可被視為XR互動、多模態多媒體網路與基礎模型／代理式AI相互耦合的整體系統，並在社群連結、內容推薦與網路效能等面向產生新的結構與動態。關鍵技術包括AI基礎模型、圖基礎模型、生成模型、知識增強生成、分析問題的NP困難度及不可近似性的方法、整數/線性/半正定規劃、動態規劃、隨機湊整、對偶理論、抽樣方法等高階演算法設計技巧。
A. Data Mining, Machine Learning, Foundation Model, and Algorithm Design for the Agentic AI Era:
Research tensor decomposition, neural network, machine learning, and other technical solutions for:
1. Virtual, augmented, and extended reality (VR/AR/XR) recommendation system (e.g., motion-sickness-aware and obstacle-avoiding path planning, multimodal foundation-model–driven content/display configuration recommendation, LLM-based context-aware agentic and proactive recommendation with knowledge augmentation (RAG), causal preference learning for recommendation, responsible-AI NFT recommendations, and data-synthesis–driven XR scenario generation for training and evaluation).
2. Social influence analysis and optimization (e.g., foundation-model–based multi-faceted influence learning, generative AI for dynamic social network modeling, multimodal analysis across virtual–physical social dynamics, causal influence learning, responsible AI (fairness-aware) influence maximization, personalized density-based group queries, graph reasoning with substructure information fusion, explainable GNN-based counterfactual learning, and agentic exploration for influence diffusion optimization).
3. Recommendation systems for other applications (e.g., LLM-augmented recommendation with knowledge augmentation (RAG), agentic-workflow multi-step personalization, explainable AI (XAI) for recommendation and mining, causal attribution of recommendations, multi-player data poisoning attacks, explainable anomaly detection in heterogeneous recommenders, Mixture-of-Experts (MoE)-enabled large-scale inference, and synthetic-data–driven attack/anomaly scenario generation and refinement for robust evaluation, group activity planning, and human-AI hybrid user recommendation).
B. Algorithm design and analysis for next-generation networks:
We analyze NP-hardness, design approximation algorithms, and use advanced algorithm techniques (e.g., approximation algorithms, competitive algorithms, and AI-based algorithms) to solve problems in next-generation networks.
1. Extended reality (XR) applications (e.g., design resource allocation and scheduling algorithms for wireless/wireline networks, select synthesized and transmitted scenes in multi-view3D videos, configure the parameters of view synthesis and cybersickness alleviation, and incorporate semantic-aware communications to optimize transmission efficiency and users' immersive experiences).
2. Low earth orbit satellite network (e.g., incorporate multicast traffic engineering, 6G, and Direct Satellite-to-Device (DS2D) communications, including UAV, satellite, and space communications, jointly consider energy efficiency and sustainability.)
3. Mobile edge computing networks (e.g., incorporate Generative Digital Twins and distributed AI architecture to build high-performance and reliable social IoT and crowdsourcing systems, leverage generative AI for missing sensor data imputation/completion, generate edge-side synthetic “virtual scenarios” (simulation-based testing) for algorithm evaluation to reduce reliance on real-world data, and validate system performance via real AI models and datasets).
4. Optimization problems for AI networking (e.g., consider AI architectures, e.g., federated learning and GNN, to design dynamic routing algorithms, choose data sources and AI training features, and control topology for minimizing the total bandwidth and computation cost and ensuring line/node capacity and service requirements).
|
其他資訊(Other Information)
|
|
主持人(PI)
吳廸融
Ti-Rong Wu
|
研究主題(Research Topic)
深度強化式學習與電腦遊戲
Deep Reinforcement Learning and Computer Games
|
研究介紹(Introduction)
深度強化式學習近年來於許多領域取得優異的成果，特別是電腦遊戲，如擊敗世界圍棋冠軍李世石的AlphaGo。本研究將探討應用各種深度強化式學習之技術於電腦遊戲上，包含但不限於：棋盤類遊戲如圍棋、五子棋、黑白棋以及電玩遊戲等。
實習生將會參與開發遊戲、使用深度強化式學習算法訓練遊戲程式以及改善搜尋演算法效能等。歡迎對深度強化式學習演算法以及電腦遊戲有興趣的同學加入。也歡迎表現良好的同學於實習後繼續與實驗室合作，參與競賽或發表論文。
Deep reinforcement learning (DRL) has achieved significant success in many fields in recent years, especially in computer games, such as AlphaGo defeating world Go champion Lee Sedol. This research study will focus on applying various DRL techniques to computer games, including but not limited to, board games and video games such as Go, Gomoku, Othello, and Atari games.
Interns will participate in developing computer games, training game-playing programs through DRL algorithms, and improving the performance of search algorithms. Students interested in DRL and computer games are welcome to join us. After the internship, students who perform well are welcome to continue to work with us, to participate in activities such as computer game tournaments or publish papers.
|
其他資訊(Other Information)
|
|
主持人(PI)
林仁俊
Jen-Chun Lin
|
研究主題(Research Topic)
連結寫實與動畫：具表現力的三維鏡頭運動編排與多維度運動補間
Bridging Realism and Animation: Expressive 3D Camera Motion Authoring and Multidimensional Motion In-betweening
|
研究介紹(Introduction)
請參考下面英文介紹:
As immersive multimedia becomes a cornerstone of next-generation storytelling, there is growing demand for tools that generate not only lifelike 3D character motion but also expressive camera movements aligned with narrative intent. Yet, two key bottlenecks continue to constrain creative freedom and scalability in 3D animation:
(a) expressive camera motion authoring, which demands expert knowledge to choreograph spatially and emotionally engaging trajectories; and
(b) multidimensional motion in-betweening, which requires synthesizing high-fidelity, coherent transitions in body pose, shape, gender, emotion, style, and species across sparse keyframes (e.g., 2D photos). This broader scope enables seamless transformations between human and non-human characters—critical for applications such as fantasy animation, creature design, and virtual performances.
Both challenges are particularly salient in long-form or multimodal content such as virtual concerts, animated films, and interactive performances—where creators must coordinate complex interactions between visuals, music, body language, and cinematic pacing. Traditional pipelines rely heavily on motion capture, professional cinematography, and iterative post-editing, making the process costly and labor-intensive, thereby limiting accessibility and scalability.
Moreover, most current AI systems either focus on character pose in-betweening—while neglecting transitions in attributes such as gender, emotion, style, and species—or concentrate on video-driven camera motion control without considering musical cues or providing intuitive user interfaces.
Bridging these two domains—realistic, editable camera motion control and semantically rich, multidimensional motion in-betweening—is critical to enabling expressive, scalable, and user-guided multimedia creation. Addressing these challenges can democratize 3D animation production, reduce reliance on expert labor, and unlock new creative possibilities in entertainment, education, and virtual performances. Therefore, this project targets three key research directions:
(a) developing a user-editable camera motion controller driven by video examples and music;
(b) building a 3D character motion in-betweening framework enabling high-fidelity, coherent transitions across pose, shape, gender, emotion, style, and species from 2D photos;
(c) developing a unified framework that integrates 3D character motion in-betweening and camera motion control to ensure coherent motion and narrative alignment.
Interns are expected to conduct research on selected topics, such as user-editable camera motion control, 3D character motion in-betweening, or related areas. After the internship, students who perform well may continue working with the laboratory on research projects and paper publications.
|
其他資訊(Other Information)
|
|
主持人(PI)
蔡懷寬
Huai-Kuang Tsai
|
研究主題(Research Topic)
用 AI 探索生物資訊
Explore Bioinformatics with AI
|
研究介紹(Introduction)
近年來，隨著高通量定序、蛋白質體技術與人工智慧方法的快速發展，生物資訊已成為解析複雜生命系統的重要核心工具。本實驗室以生物資訊與人工智慧為研究基礎，結合來自國內外研究機構的大尺度生物資料，致力於從資料中理解基因體與蛋白質層級的調控機制與結構特徵，並探索其在疾病、逆境與演化中的角色。
本實驗室的研究重點在於以資訊思維與 AI 驅動生物問題的探索，結合機器學習、深度學習與多體學資料，應用於蛋白質、DNA 與 RNA 序列的優化設計、基因體結構分析、藥物標靶親和性預測，以及暗基因（功能未知基因）的系統性解析。整體而言，我們關注的是如何整合多體學資料、基因體結構、網路分析與機器學習方法，從系統層級理解複雜的生物問題，並將資料轉化為可解釋的生物知識。
本實驗室誠摯歡迎對生物資料分析與跨領域研究有興趣的大專生申請暑期實習。申請者可來自生物、生命科學、資訊工程、電機、數學或相關背景，並應熟悉至少一種程式語言（如 Python、R、C/C++），且對生物問題保持好奇心。實習期間將提供生物資訊與資料分析的基礎訓練，並依學生背景安排實際研究題目，讓學生參與真實研究資料的分析與討論。若你希望在暑期深入了解 AI 與計算方法如何應用於現代生物研究，誠摯歡迎加入我們的研究團隊。
In recent years, advances in high-throughput sequencing, proteomics, and artificial intelligence have made bioinformatics a central tool for understanding complex biological systems. Our laboratory is grounded in bioinformatics and artificial intelligence, integrating large-scale biological data from domestic and international research institutions. We aim to investigate regulatory mechanisms and structural features at the genomic and proteomic levels, and to explore their roles in disease, stress responses, and evolution.
Our research focuses on AI- and information-driven exploration of biological problems, combining machine learning, deep learning, and multi-omics data. These approaches are applied to the optimization and design of protein, DNA, and RNA sequences, genome structural analysis, prediction of drug–target binding affinity, and the systematic characterization of dark genes (genes with unknown or poorly characterized functions). More broadly, we aim to integrate multi-omics data, genome structure, network analysis, and machine learning methods to achieve systems-level understanding and interpretable biological insights.
We warmly welcome undergraduate students with strong interests in biological data analysis and interdisciplinary research to apply for our summer internship program. Applicants may come from backgrounds in biology, life sciences, computer science, electrical engineering, mathematics, or related fields, but should be familiar with at least one programming language (e.g., Python, R, or C/C++) and have genuine curiosity about biological questions. During the internship, students will receive foundational training in bioinformatics and data analysis, and will participate in hands-on research projects using real biological datasets. If you are interested in learning how AI and computational approaches are applied to modern biological research, we warmly invite you to join our research team.
|
其他資訊(Other Information)
|
|
主持人(PI)
楊柏因
Bo-Yin Yang
|
研究主題(Research Topic)
後量子密碼學
post-quantum cryptography
|
研究介紹(Introduction)
後量子密碼學，即在量子運算時代依然存在的非對稱密碼學，的實例化與實現。
The instantiation and implementation of post-quantum cryptography, asymmetric cryptography that survives quantum computing
|
其他資訊(Other Information)
|
|
主持人(PI)
修丕承
Pi-Cheng Hsiu
|
研究主題(Research Topic)
可持續的微型機器學習
Sustainable TinyML
|
研究介紹(Introduction)
此計畫屬於嵌入式系統研究領域，特別關注「可持續的微型機器學習」，在促進邊緣智能發展的同時兼顧環境永續。我們開發系統軟體，以協助人工智慧研究人員輕鬆部署並高效執行他們的深度學習模型在配有微控制器的微型裝置上。學生將整合並應用我們開發的「深度學習推論引擎」與「類神經網絡架構搜尋工具」於超低功率嵌入式裝置，並學習到系統實作與開發的經驗。
This project's scope lies in the area of embedded systems, with a special focus on enabling tiny devices to execute deep neural networks (DNN) in an environmentally sustainable manner. We develop system software for AI researchers to easily deploy and efficiently execute their DNN models on tiny devices that feature microcontrollers. You are expected to gain rich hands-on experience in prototype implementations and hacking system kernels by integrating and applying our previously developed deep learning inference engine and neural architecture search tool to ultra-low power embedded platforms.
|
其他資訊(Other Information)
|
|
主持人(PI)
張佑榕
Ronald Y. Chang
|
研究主題(Research Topic)
AI/ML賦能的第六代無線通訊
AI/ML-Empowered 6G Wireless Communications
|
研究介紹(Introduction)
見英文介紹
INTERNS will explore the intersection of AI/ML and sixth-generation (6G) wireless communications, especially deep learning for mmWave, multiple-input multiple-output (MIMO), beamforming, beam management, cell-free networks, non-terrestrial networks (NTNs), integrated sensing and communication (ISAC), etc.
INTERNS will participate in weekly and ad-hoc meetings, conduct research, and prepare research reports/slides/presentations/research papers. Paid extensions after the official two-month period are available upon demonstration of satisfactory performance and mutual agreement.
EXPERIENCE with Wireless Communication (communication systems, signals and systems, digital signal processing, etc.) and Machine Learning (convolutional neural networks (CNNs), graph neural networks (GNNs), reinforcement learning (RL), federated learning (FL), etc.) is a plus.
// Lab areas: Wireless Communication and Networking, Machine Learning for Wireless Communications
// Lab former members' first positions: Ph.D. students at Purdue, UC Davis, OSU, USC, etc.; M.S. student at Stanford, USC, UCSD, CMU, TU Munich, etc.; senior engineers at MediaTek, Realtek, etc.
|
其他資訊(Other Information)
|
|
主持人(PI)
王釧茹
Chuan-Ju Wang
|
研究主題(Research Topic)
產業與金融應用之大型語言模型代理式（Agentic）框架研究
Agentic Large Language Model Frameworks for Industrial and Financial Applications
|
研究介紹(Introduction)
本研究旨在探討並開發應用於產業與金融場景之大型語言模型代理式（agentic）系統框架。研究將聚焦於具備規劃能力、工具調用、檢索整合與多步推理能力的 LLM 代理，使其能在真實世界環境中執行複雜任務並輔助決策。代理系統將結合檢索增強生成（Retrieval-Augmented Generation, RAG）、外部知識庫與領域專用工具，以提升系統的可靠性、可解釋性與實用性。
應用面向可能涵蓋多個領域，例如：基於非結構化金融資料的代理式金融分析與問答系統，或用於產業場景的智慧代理，如決策支援、系統監控與特定用途程式撰寫輔助等。研究亦將關注代理行為的評估、系統化實驗設計，以及代理與外部系統互動的最佳實務。
除模型與系統設計外，研究參與者將能實際參與真實世界資料處理，學習在 Unix-like 系統中進行大規模資料管理與實驗，並透過前端網頁技術將代理決策與系統輸出進行視覺化展示。
This research aims to investigate and develop agentic large language model (LLM) frameworks for real-world industrial and financial applications. The focus is on designing LLM-based agents with capabilities such as planning, tool usage, retrieval integration, and multi-step reasoning, enabling them to perform complex tasks and support decision-making in practical environments. These agentic systems will integrate retrieval-augmented generation (RAG), external knowledge sources, and domain-specific tools to improve reliability, interpretability, and usability.
The research may span multiple application domains. Examples include agent-based financial analysis and question-answering systems operating on large-scale unstructured financial data, as well as industrial AI agents for tasks such as decision support, system monitoring, or code assistance for specialized programming languages. Particular attention will be paid to agent behavior analysis, systematic evaluation, and effective interaction between agents and external systems.
Beyond model and system design, participants will gain hands-on experience with real-world data, learn to manage large-scale data and conduct structured experiments in Unix-like environments, and acquire skills in visualizing agent decisions and system outputs through front-end web programming.
|
其他資訊(Other Information)
|
|
主持人(PI)
王柏堯
Bow-Yaw Wang
|
研究主題(Research Topic)
密碼程式形式化驗證
Formal verificaiton of cryptographic programs
|
研究介紹(Introduction)
本研究將開發編譯器中間表示式轉換至形式驗證語言，以利程式設計師驗證所開發之密碼程式。
This research plans to transform GCC or LLVM Intermediate Representations to the formal verification CryptoLine to help C/C++/Java/Rust programmers verify their cryptographic programs.
|
其他資訊(Other Information)
|
|
主持人(PI)
陳駿丞
Jun-Cheng Chen
|
研究主題(Research Topic)
基於視覺生成式AI之空間智慧與安全應用探索
Exploring Spatial Intelligence and Security Applications Based on Visual Generative AI
|
研究介紹(Introduction)
我們的實習專題聚焦於視覺生成式 AI 與空間智慧，涵蓋擴散模型(如diffusion model, flow matching, etc)、自迴模型以及多模態大型語言模型等方向，目標應用包括文字到影像生成與編輯、連續與細緻可控生成（fine-grained controllability），以及三維與四維生成式世界模型（generative world models) 研究保持幾何／時間一致性、長時序連貫性與互動式生成。同時，我們也研究多模態大型語言模型的 AI 安全議題，涵蓋提示注入（prompt injection）、對抗式提示（adversarial prompts），以及更廣泛針對大型語言模型或多模態大型語言模型驅動的代理式 AI 系統（agentic AI systems）之攻防與防禦。
表現優秀的實習生在實習結束後，會獲得繼續擔任兼任研究助理的研究機會，並可將成果繼續精進並整理成論文，投稿至 CVPR、ICCV、ECCV、ACM MM、NeurIPS 等頂尖國際會議。
Our intern projects focus on visual generative AI and spatial intelligence—including diffusion, flow matching, autoregressive models, and multimodal large language model—for text-to-image generation and editing, fine-grained controllability, and 3D/4D generative world models (e.g., geometric/temporal consistency, long-horizon coherence, and interactive generation). In parallel, we study AI security for multimodal LLMs, covering prompt injection, adversarial prompts, and broader attack/defense settings for LLM/MLLM-based agentic AI systems. High-performing interns may be invited to continue as part-time research assistants after the internship, and we actively support turning successful projects to top-tier conferences such as CVPR, ICCV, ECCV, ACM MM, NeurIPS, etc.
|
其他資訊(Other Information)
|
|
主持人(PI)
陳亮廷
Liang-Ting Chen
|
研究主題(Research Topic)
程式語言與數學基礎—理論與應用
Programming language and foundations of mathematics – its theory and applications
|
研究介紹(Introduction)
此暑期實習計劃將著重於程式語言理論的相關工具使用、問題解決的能力以及基礎理論理解。
# 背景說明
程式設計與建構式數學本質上兩者互相呼應。以型別論（type theory）為基礎的程式語言可用來論證數學事實。而具備計算意義的數學基礎（foundation of mathematics）可用作程式語言以及其邏輯語言。我們可以將邏輯敘述看作是程式的型別，將證明看作是程式，將證明檢查的過程看作是型別檢查等的聯繫。此邏輯與計算之間的聯繫泛稱為 Curry-Howard 對稱。在此框架下，我們可以將構造式數學的定理視作程式，例如「所有自然數必然是偶數或是奇數」之證明，可視作為給定任一自然數，構造出附帶奇偶證明的程式計算。
# 實習主題
將採用同時具備依值型別（dependently typed language）及可用作互動式定理證明（interactive theorem prover）的程式語言 Agda 探索計算機科學中的計算與邏輯連結。實習規劃上有應用與理論兩大方向可選：
1. 應用層面包括探討如何使用 Agda 建構保證正確（correct by construction）的函數式程式（functional program）或探索數學結構或理論的計算性質。
2. 理論方向則是了解探索如何用數學模型設計並驗證定理證明器的核心系統，並用 Agda 建構驗證其數學模型。此方向會需要範疇論或是高等數學背景，對此領域有興趣者建議至少修過若干數學系課程。
過往實習題目可參考網站：https://l-tchen.github.io
# 申請條件
此研究實習計劃需進行嚴格數學論證，以及針對個人學術興趣發展適合的題目。
因此欲申請此暑期實習計劃，請在申請信中針對：「數學成熟度」以及自己的「研究興趣」（任何主題皆可）說明。
# 實習時程規劃
第一到二周：參與在台灣大學舉辦的「邏輯、語言與計算暑期研習營」簡稱 FLOLAC (https://flolac.iis.sinica.edu.tw)。
第三到五周：共同學習 Agda 的使用，並在此段時間根據研究興趣共同構思研究方向主題以及問題。
第六到九周：根據研究問題閱讀相關文獻、探索、並且動手解決，並且提供學術寫作上的指導。
註：請依個人狀況自行選擇正式修課或旁聽參與 FLOLAC 課程，但作為實習狀況評估皆須參與考試。
若您申請 FLOLAC 外地實習（詳情請見 FLOLAC 網站 https://flolac.iis.sinica.edu.tw），
請在自傳內特別註明，並提出您適合研究此主題的具體理由，方能優先錄取。
This summer internship programme will focus on the use of programming language theory, problem-solving skills, and understanding of theories.
# Background
Programming and constructive mathematics are essentially interconnected. Programming languages based on type theory can be used to reason about mathematical facts. The computational foundation of mathematics can be used as programming languages and their logical system. We can view logical statements as types of programs, proofs as programs, and the proof-checking process as type checking, etc. This connection between logic and computation is generally referred to as Curry-Howard correspondence.
# Topic
The dependently typed language and the interactive theorem prover Agda will be used to explore the computational and logical characteristics of programming languages. There are two main internship directions to choose from:
Application: Program verification or constructive foundations of mathematics
Theory: Mathematical models of type theory
# Requirements
This research internship program requires extensive and rigorous mathematical proofs, as well as the development of suitable topics tailored to individual academic interests.
Therefore, to apply for this summer internship program, please specify your "mathematical maturity" and "research interests" (any topic is acceptable) in your application letter.
# Schedule
Weeks 1-2: Participation in the "Summer Workshop on Logic, Languages, and Computation" (FLOLAC) held at National Taiwan University (https://flolac.iis.sinica.edu.tw).
Weeks 3-5: Collaborative learning of Agda, and during this period, jointly conceiving research directions and questions based on research interests.
Weeks 6-9: Reading relevant literature, exploring, and solving problems based on research questions, with guidance provided on academic writing.
Note: Please choose to formally enrol in or audit FLOLAC courses based on your individual circumstances, but all will require an exam for evaluation.
If you are applying for a FLOLAC field internship (details can be found on the FLOLAC website: https://flolac.iis.sinica.edu.tw),
please specifically state this in your personal statement and provide concrete reasons why you are well-suited to research this topic; this will give you priority in admission.
|
其他資訊(Other Information)
|
|
主持人(PI)
蘇黎
Li Su
|
研究主題(Research Topic)
音樂分析檢索與計算音樂學
Music Information Retrieval and Computational Musicology
|
研究介紹(Introduction)
音樂與文化科技實驗室(Music and Computational Thinking Lab）成立於2017年，致力於音樂資訊檢索（Music Information Retrieval, MIR）與計算音樂學（Computational Musicology）的研究，關注如何運用人工智慧與資料分析方法，協助機器理解、分析音樂，並與人類互動。
在音樂資訊檢索方面，我們的目標包括建立能夠進行完整音樂理解的模型，以及發展具備即時反應能力的音樂追蹤與互動系統。近年的研究聚焦於自監督學習(self-supervised learning)與基礎模型(foundational model)在音樂分析任務中的應用，涵蓋特定樂器或曲風導向的音樂採譜問題（如旋律、吉他獨奏與小提琴指法辨識等），以及結合樂譜與音訊的跟譜（score following）與自動伴奏。我們亦關注多模態音樂處理，探討音樂與肢體動作、表演行為之間的關係，並進一步發展符號音樂理解方法，如功能和聲辨識與聲部抽取，期望透過現代深度學習模型提升對音樂結構與語意的掌握能力。
在計算音樂學方面，MCTL 的長期願景是以 MIR 技術支持新的音樂學研究典範。我們關心的議題包括但不限於音樂家合作網絡的重建、古典音樂作品的典範化（canonicalization），以及世界音樂中隱含之音樂理論與知識結構的探索。相關研究多與國內外音樂學、人文學科與資訊領域學者合作，結合資料驅動方法與音樂學詮釋，嘗試回應實際音樂研究中的核心問題。
MCTL 歡迎對音樂、人工智慧、資料分析或跨領域研究有興趣的大學生申請實驗室實習。我們特別重視學生的學習動機與探索精神，並提供參與真實研究計畫、資料建構與論文研究的機會。
The Music and Computational Thinking Lab (MCTL), established in 2017, focuses on research in Music Information Retrieval (MIR) and Computational Musicology. Our work explores how artificial intelligence and data-driven methods can be used to enable machines to understand, analyze, and interact with music in meaningful ways.
In the area of music information retrieval, our research aims to develop computational models for comprehensive music understanding as well as real-time music tracking and interactive systems. In recent years, we have focused on applying self-supervised learning and foundation models to music analysis tasks. Our projects include instrument- or genre-specific music transcription—such as melody extraction, guitar solo transcription, and violin fingering recognition—as well as score following and automatic accompaniment that integrate symbolic scores with audio signals. We also investigate multimodal music processing, studying the relationship between music, body movement, and performance behavior. In addition, we develop methods for symbolic music understanding, including functional harmony recognition and voice extraction, with the goal of enhancing the modeling of musical structure and semantics using modern deep learning techniques.
In computational musicology, MCTL’s long-term vision is to advance new paradigms of music research supported by MIR technologies. Our research interests include reconstructing collaborative networks of musicians, the canonicalization of classical music repertoires, and the discovery of musical theories and knowledge embedded in world music traditions. These studies are often conducted in collaboration with scholars in musicology, the humanities, and information science at universities in Taiwan and abroad, combining data-driven approaches with musicological interpretation to address fundamental questions in music research.
MCTL welcomes undergraduate students who are interested in music, artificial intelligence, data analysis, or interdisciplinary research to apply for research internships. We value strong motivation and intellectual curiosity and offer opportunities to participate in real-world research projects, dataset development, and academic publication.
|
其他資訊(Other Information)
|
|
主持人(PI)
柯向上
Hsiang-Shang ‘Josh’ Ko
|
研究主題(Research Topic)
函式程式與型式化證明（但不要太多）
Functional programs with formal proofs (but not too many)
|
研究介紹(Introduction)
每當我們寫出數學定理和證明，其實也就寫出具有依值型別（dependent type）的函式程式（functional program）。從一開始 Curry、Howard 等人察覺到幾套獨立發明的數理邏輯系統和計算系統竟有相同本質，到 Martin-Löf 發明 Type Theory 作為數學和程式寫作的大一統基礎，隨後衍生出眾多證明輔助器（proof assistants）和依值型別程式語言（dependently typed programming languages），我們現在已能以同一型式編寫程式和正確性證明。然而此類型式化證明成本相當高昂，減少型式化證明負擔一直是程式語言與型式驗證研究領域的主要挑戰之一。
實習的焦點會放在 Agda 這個程式語言學界常用的語言，先學習基本的依值型別程式寫作（dependently typed programming），再依興趣往幾個方向延伸：
1. 將更多演算法與資料結構改寫為依值型別程式，追求讓證明與程式合而為一，從而不需寫太多額外證明；與證明結合的程式也往往比原先版本更有條理結構，能帶來更清晰甚至新穎的理解。
2. 依值型別程式已發展出一些設計樣式，但現況是這些設計樣式得靠程式撰寫者手工套用在不同程式上；一個研究方向是整理這些設計樣式，再以泛資料型別技術（datatype-generic programming）打造為可重用的程式庫組件（相當於數學上將非正式的證明直覺、技巧寫為正式定理）。
3. 探究依值型別程式的語言機制和設計樣式如何減省證明（與不使用這些機制和樣式的狀況對照），並發展理論協助程式撰寫者分析何時適合／不適合使用這些機制和樣式（類似於數學上解題時有些定理和法則可試著套用），以及啟發新的證明減省技術。
實習型式預設類似讀書會，自行研讀材料和動手寫 Agda 程式，並與老師同學們分享討論；若有餘裕也可做個小專案。若對方向、題目、型式有其他想法，亦可和老師討論。前兩週須參加於台大舉辦的「邏輯、語言、與計算」暑期研習營（https://flolac.iis.sinica.edu.tw；修課或旁聽均可），暫定 6/29 開課，除非已獲得偶數年修業證書，或有其他特殊情況請儘早先和老師商量。
申請材料內請務必敘述對此主題有興趣的理由以及想達成的（大致）目標。
參考讀物請見英文版末段。
Whenever we write down a mathematical theorem and its proof, we have also written down a dependently typed functional program. Based on the Curry–Howard correspondence, which stemmed from observations that several independently invented logical and computational systems were nevertheless essentially the same, Martin-Löf developed Type Theory as a unified foundation for mathematics and programming, which has spawned numerous proof assistants and dependently typed programming languages where we can write programs and their correctness proofs in the same unified form. However, the cost of writing such formal proofs is exceedingly high, and reducing the burden of formal proofs has been a major challenge in the research area of programming languages and formal verification.
We will focus on Agda, which is a popular language in the programming languages research community. We will start by learning basic dependently typed programming, and then there will be a few possible directions to explore:
1. Rewrite more algorithms and data structures into dependently typed programs, with the aim of fusing proofs into programs and avoiding separate proofs; in addition, such programs usually have clearer structure that brings better and even new understanding.
2. There are design patterns in dependently typed programming, but usually the programmer has to manually apply them to individual programs; one research direction is to collect these patterns and then turn them into reusable library components using datatype-generic programming (cf formulating informal proof intuitions and techniques as formal theorems in mathematics).
3. Investigate how dependently typed language mechanisms and design patterns reduce proofs (by comparing with programs that do not use these mechanisms and patterns), and develop theories that help the programmer to analyse when (not) to use these mechanisms and patterns (cf trying to apply theorems and rules when solving mathematical problems), and stimulate the development of new proof-reducing mechanisms in the long term.
The default format will be like a study group, where each member will study relevant materials, write Agda programs, and share their findings and discuss with the group. If time permits, there is also an opportunity to undertake a small project. If there are other ideas about direction, topic, or format, please feel free to discuss them with the supervisor. You will be required to attend the Formosan Summer School on Logic, Language, and Computation (FLOLAC) at the National Taiwan University during the first two weeks (https://flolac.iis.sinica.edu.tw, taught in Mandarin; either signing up for credits or just sitting in), starting from 29 June. Exceptions apply if you have already obtained a certificate in an even-numbered year or if special circumstances arise, and these should be discussed with the supervisor as early as possible.
In your application, please state why you are interested in this research topic and (roughly) what you want to achieve.
References
* Ingo Blechschmidt [2025]. Let’s play Agda: Running abstract mathematical proofs as programs. https://lets-play-agda.quasicoherent.io
* Hsiang-Shang Ko, Shin-Cheng Mu, and Jeremy Gibbons [2024]. Binomial tabulation: A short story. https://arxiv.org/abs/2503.04001
* Hsiang-Shang Ko and Shin-Cheng Mu [2025]. Bottom-up computation using trees of sublists: A dependently typed approach. https://github.com/josh-hs-ko/BT/raw/main/JFP/BT.pdf
* Hsiang-Shang Ko [2021]. Programming metamorphic algorithms: An experiment in type-driven algorithm design. The Art, Science, and Engineering of Programming, 5(2):7:1–34. https://doi.org/10.22152/programming-journal.org/2021/5/7
* Hsiang-Shang Ko [2025]. Specifying queue order using parametric types, concretely. https://josh-hs-ko.github.io/blog/0047/
* Hsiang-Shang Ko [2026]. Specifying queue order using parametric types, abstractly. https://josh-hs-ko.github.io/blog/0048/
* Conor McBride [2011]. Ornamental algebras, algebraic ornaments. https://personal.cis.strath.ac.uk/conor.mcbride/pub/OAAO/LitOrn.pdf
* Hsiang-Shang Ko, Liang-Ting Chen, and Tzu-Chi Lin [2022]. Datatype-generic programming meets elaborator reflection. Proceedings of the ACM on Programming Languages, 6(ICFP):98:1–29. https://doi.org/10.1145/3547629
|
其他資訊(Other Information)
|
|
主持人(PI)
鐘楷閔
Kai-Min Chung
|
研究主題(Research Topic)
量子/古典密碼學、複雜度理論或量子演算法之獨立研究
Independent Research on Quantum/Classical Cryptography, Complexity Theory, or Quantum Algorithm
|
研究介紹(Introduction)
The intern is expected to perform independent research on selected topics in Quantum/Classical Cryptography, Quantum/Classical Complexity Theory, Quantum Algorithms, or general theoretical computer science (TCS) that interest him/her. This often starts by surveying research papers and presenting them to the PI. Along the way, the intern can identify research questions with the PI, perform independent study, and discuss them with the PI in research meetings.
Students interested in theoretical computer science, particularly on the abovementioned topics, are encouraged to apply.
Please elaborate on your interests in TCS in your application.
The intern is expected to perform independent research on selected topics in Quantum/Classical Cryptography, Quantum/Classical Complexity Theory, Quantum Algorithms, or general theoretical computer science (TCS) that interest him/her. This often starts by surveying research papers and presenting them to the PI. Along the way, the intern can identify research questions with the PI, perform independent study, and discuss them with the PI in research meetings.
Students interested in theoretical computer science, particularly on the abovementioned topics, are encouraged to apply.
Please elaborate on your interests in TCS in your application.
|
其他資訊(Other Information)
|
|
主持人(PI)
穆信成
Shin-Cheng Mu
|
研究主題(Research Topic)
函數語言與命令語言程式之正確性推理
Reasoning about Functional and Imperative Programs
|
研究介紹(Introduction)
我的研究興趣是程式語言與函數程式設計(functional programming)，近年來也包括命令式語言(imperative languages)的推理，與 concurrent 程式的型別系統與。它們的共同點是使用符號推理的方式確保程式的正確性。不論在哪個典範中，我們都希望把「寫程式」視作一個可用數學與邏輯方式推理的行為。程式的正確性可用型別系統或邏輯推演保證，甚至可用規格與需求開始，經由數學方法一步步推導出程式。
本領域可做的大方向包括
* 程式推導(program derivation)與證明。「程式推導」是由一個問題的描述（通常也是個函數語言程式）開始，一步步用數學方法將解決該問題的演算法推衍出來的技術。我們可能會挑一些有趣的演算法問題，試著找出他們的解法。
* 純函數資料結構與二進位數演算的關係：有些資料結構（如 binomial heap, Okasaki 的 random access list 等等）和二進位數的表示與運算高度相關。藉由依值(dependent-type)型別系統的協助，我們是否能由二進位數的遞增、加法等等函數推導出資料結構上的相對應操作？
* 以函數語言為工具，開發 Hoare logic 與命令式語言程式推導使用的教學系統。我們開發了一個協助程式推導的整合環境 Guabao (https://scmlab.github.io/guabao/ ), 參與此計畫可接觸許多程式語言實作相關的技術 --- 不僅是編譯(compilation)。
* 組合語言程式的正確性證明：該怎麼做、掌握什麼原則？
* 設計幫助推理用的符號、程式語言、型別系統等。
* 研究 concurrent 程式以及其型別系統 (session type) 與邏輯之關係。
如對以上題目有興趣，在三個月的實習期間，我們可用一到一個半月的時間學習相關理論（函數編程、型別、邏輯等），用剩下的時間研究新東西或開發系統。
My research interest concerns programming language and functional programming, and extends to Hoare logic and type systems for concurrent programs. The common theme is that programming is seen as a formal, mathematical activity. Correctness of a program can be guaranteed by logical reasoning or type system. Or, a program can even be derived stepwise from its specification.
Possible topics include:
* Program derivation and reasoning. Program derivation is the technique of constructing an algorithm from a problem specification (usually also a functional program), in a step-wise manner such that every step is mathematically valid. We may pick some interesting algorithmic problems, and try to derive programs that solve them.
* Data structures based on binary number representations. Some data structures, such as binomial heap, random access list of Okasaki, etc, are closely related to representation of binary numbers. Can we derive operations on such data structures from the corresponding operations on binary numbers, with the help of dependent types?
* Develop tools for reasoning about imperative programs, using a functional programming language. We have developed an integrated environment, Guabao (https://scmlab.github.io/guabao/ ), about which there is still plenty of theory and implementation to be done. One may learn plenty of techniques related to programming language implementation --- which is way more than compilation.
* Design symbols, languages, or type systems that aids the programmers in reasoning about programs;
* Study the type system (session type) for concurrent programs and its relationship with logic;
More details can be discussed. If you are interested, we can spend the first 1 to 1.5 months of the internship studying the background knowledge, before diving into developing something new.
|
其他資訊(Other Information)
|
|
主持人(PI)
吳真貞
Jan-Jan Wu
|
研究主題(Research Topic)
高效能深度學習計算於異質多處理器環境之資源配置與排程技術
Resource Allocation and Scheduling for Improving Deep Learning Execution Time on Heterogeneous Multi-processor Platforms
|
研究介紹(Introduction)
將多個網絡組合成混合模型或多模型是提高 DNN 性能的可行方法。這些模型可以
通過利用不同網絡的優勢來解決更複雜的任務。 例如，多車型的應用包括自動駕
駛汽車和語音助手。另一方面，異質系統架構在現代計算機中被廣泛採用。它混合
了各種類型的計算設備，可更有效地利用資源並提高多種工作負載的效能。例如
，谷歌雲服務器可能包含許多 CPU、GPU 和 TPU. 如果可以有效地利用系統資源
，異質系統架構將可提高 DNN 的計算效能。然而，TensorFlow、PyTorch 和
TVM 等現代深度學習平台主要是為同質系統設計的。他們只在一種類型的設備上
運行 DNN。此外，這些平台也不支援混合模型和多模型。
為了解決這些問題，本計畫將發展可在異質多處理器環境中支援高效能且自動化的
混合型/多模型的深度學習計算系統。神經網絡可以表示為計算圖。 問題變成如何
將圖形映射到異質計算設備。 本計劃將分兩個階段解決此類映射問題：(1) 資源分
配階段將圖節點分配給設備，(2) 排程階段確定圖節點的執行順序。我們針對此二
階段映射問題提出數種高效率的演算法及系統實作。本計畫特色在於充分發揮各階
層的平行度,包括 data parallelism, pipeline parallelism(例如,跨設備切割模型，工
作負載以管道方式流經拆分的子模型),以及tensor parallelism(例如，AI 加速器使用
VLIW 來同時計算許多向量或矩陣)
Because of the demand for higher prediction accuracy, today’s neural networks
are becoming deeper, wider, and more complex, typically with many layers and a
large number of parameters. Moreover, combining multiple networks into a
hybrid- or multi-model is a viable way to improve the performance of DNNs.
These models can resolve more complex missions by leveraging the strengths of
different networks. On the other hand, heterogeneous system architectures
(HSAs) are getting widely adopted in modern computers. It mixes various types
of computing devices and communication technologies, allows for more efficient
use of resources and improved performance for many types of workloads. Such
HSAs provide ample opportunity to improve the performance of DNNs if the
system resources can be efficiently and effectively utilized.
However, modern deep learning platforms such as TensorFlow, PyTorch, TVM,
etc. are mainly designed for homogeneous systems. They run DNNs only on one
type of devices, leaving other devices of the heterogeneous systems unused.
Furthermore, hybrid- and multi-models are overlooked in these platforms. Hence,
developers need to manually tune the performance on the target hardware, which
usually needs expert knowledge and experience.
To address these issues, we will design a runtime system to handle the execution
of hybrid-/multi-models on HSAs efficiently and automatically. A neural network
can be represented as a computational graph. The problem becomes how to map
the graph(s) to the heterogeneous devices. We plan to tackle such mapping
problem in two phases: (1) the resource allocation phase assigns graph nodes to
devices, and (2) the scheduling phase determines the execution order of the
graph nodes. Three core issues will be addressed in resource allocation: (1) We
need to assign operations to appropriate computing devices to minimize the
computation cost. (2) We need to assign the operations so that no operations use
the same computing device at the same time. (3) We must choose the
appropriate communication medium when two related operations are mapped to
different computing devices, so as to reduce the communication overhead.
The challenge in designing an efficient scheduling is how to exploit the
parallelism among the computing devices while retaining data dependency. We
consider three types of parallelism: data parallelism (DP), pipeline parallelism
(PP), and tensor parallelism (TP). DP is a widely adopted technique of dividing a
large workload into smaller subsets and executing multiple copies of the neural
network on these subsets simultaneously on the devices. PP divides the model
across the devices and workload flows through the split sub-models in a pipeline
manner. It can be useful for training very large or complex models and speed up
streaming applications. TP divides the computation of a single layer across the
devices, which process different parts of the tensors in parallel. For example, the
AI accelerators (e.g., Google’s EdgeTPU) employ VLIW to simultaneously
compute many vectors or matrices. The above three parallelisms impose different
constraints and resource requirements of the devices. Therefore, a sophisticated
method is required to determine the best parallelism configuration to run the
DNNs.
|
其他資訊(Other Information)
|
|
主持人(PI)
古倫維
Lun-Wei Ku
|
研究主題(Research Topic)
(1) 創造力多模態語言模型 (2) 語言模型與新聞素養 (3) 運動科技-智慧教練 (4) 語言模型的人性模擬
(1) Multimodal LLMs for Creativity (2) LLMs and News Literacy (3) SportTech - AI Coach (4) Human-like Simulation of LLMs
|
研究介紹(Introduction)
在這些研究主題中，將學習到自然語言處理之資訊擷取、文章分類、文字生成、知識庫使用、圖像文字結合、大型語言模型等概念，另涵蓋自然語言基礎工具的使用及機器學習、深度學習的模型建立等先進技術，可與老師討論希望選擇的研究主題。實習期間會專注於上述研究主題並參與模型開發及論文撰寫。各主題研究內容詳述如下：
(1) 在創造力多模態語言模型中，我們注重在圖像與文字交匯所能帶來的創造力。相關技術可應用於圖片生成中。
(2) 語言模型與新聞素養中，我們著重於將過去一連串打擊假新聞的技術應用於教育，除了學習思辨邏輯並提高新聞素養外，也研究相關技術如何協助大型語言模型的推理能力。
(3)運動科技-智慧教練中，我們希望開發對特定運動姿態的小樣本或無樣本學習模型，並經由圖像文字結合技術，自動生成智慧教練指導語。此研究目標為真正可用的系統。
(4) 語言模型的人性模擬中，我們研究如何將語言模型的表現極度接近真實世界的真人。
實驗室尚有其他研究主題正在進行，可到
http://www.lunweiku.com/ 參考相關論文。
實習結束後，表現優良的同學可繼續與實驗室合作研究並發表論文。
Interns will learn how to use basic natural language processing tools, extract information from texts, classify documents, generate dialogs and large language model basics. Machine learning and deep learning technologies for NLP will be touched. Interns can select the topic/team they wish to join.
(1) In multimodal language models for creativity, we focus on the creativity that emerges at the intersection of images and text. Related technologies can be applied to image generation.
(2) In the intersection of language models and news literacy, we emphasize applying a series of past techniques for combating fake news to education. In addition to fostering critical thinking and enhancing news literacy, we also explore how these technologies can assist large language models in reasoning.
(3) In sports technology – smart coaching, we aim to develop small-sample or zero-shot learning models for specific sports movements. By integrating image and text technologies, we seek to automatically generate smart coaching instructions. The goal of this research is to create a truly practical system.
(4) In the human-like simulation of language models, we study how to make the performance of language models closely approximate real-world human behavior.
The lab is also conducting other research topics. For more details, you can refer to relevant papers at
http://www.lunweiku.com/.
After completing the internship, students with outstanding performance may continue to collaborate with the lab on research and publish papers.
|
其他資訊(Other Information)
|
|
主持人(PI)
李政池
Gen-Cher Lee
|
研究主題(Research Topic)
加密行動應用程式設計及AI幾器人訊息傳輸保護
Design of Encrypted Mobile Applications and Protection of AI Robot Message Transmission
|
研究介紹(Introduction)
研究Android/iOS行動裝置的多模式程式開發，並運用安全加密技術以保護AI機器人的訊息傳輸。參與此計劃使用到的相關程式語言包括C/C++/Python/Java/Kotlin/Swift/Dart/Rust，並可演練功能上線。
Research on multi-modal application development for Android/iOS mobile devices, utilizing secure encryption technology to protect AI robot message transmission. The programming languages involved in this project include C, C++, Python, Java, Kotlin, Swift, Dart, or Rust, with opportunities to practice feature deployment.
|
其他資訊(Other Information)
|
|
主持人(PI)
劉庭祿
Tyng-Luh Liu
|
研究主題(Research Topic)
Computer Vision, Artificial Intelligence, and Machine Learning
Computer Vision, Artificial Intelligence, and Machine Learning
|
研究介紹(Introduction)
Join our lab for research in Computer Vision, Artificial Intelligence, and Machine Learning. Our current focus is on developing deep learning architectures and algorithms for real-world computer vision applications.
Research areas include: (1) Generative Computer Vision Techniques; (2) Image and Video Anomaly Detection and Localization; (3) Computer Vision for 3D Point Clouds; (4) Video Action and Activity Recognition; (5) Multimodal / Vision–Language Foundation Model–Inspired Computer Vision; (6) Self-Supervised Learning for Computer Vision Applications; (7) Deepfake Detection and Security; Federated Learning Algorithms and Optimization.
Join our lab for research in Computer Vision, Artificial Intelligence, and Machine Learning. Our current focus is on developing deep learning architectures and algorithms for real-world computer vision applications.
Research areas include: (1) Generative Computer Vision Techniques; (2) Image and Video Anomaly Detection and Localization; (3) Computer Vision for 3D Point Clouds; (4) Video Action and Activity Recognition; (5) Multimodal / Vision–Language Foundation Model–Inspired Computer Vision; (6) Self-Supervised Learning for Computer Vision Applications; (7) Deepfake Detection and Security; Federated Learning Algorithms and Optimization.
|
其他資訊(Other Information)
|
|
主持人(PI)
徐讚昇
Tsan-sheng Hsu
|
研究主題(Research Topic)
人機共作式殘局知識萃取
A hybrid method for abstracting knowledge in endgames
|
研究介紹(Introduction)
對古典對局之人機共作式殘局知識萃取
We will explore knowledge embedded in large endgames databases for classical board games, such as EWN and Chinese dark chess (CDC) using a hybrid method of human and deep learning co-investigation.
|
其他資訊(Other Information)
|
|
主持人(PI)
連云暄
Yun-Hsuan Lien
|
研究主題(Research Topic)
LeRobot 實體機器人控制
Real-Robot Control with LeRobot
|
研究介紹(Introduction)
[LeRobot 實體機器人控制] 暑期實習將帶領實習生使用 LeRobot 打造一套完整的機器人學習流程，從資料收集、行為學習訓練到真實機器人上的部署。實習生會在至少 1–2 個操作任務上進行 benchmark、錄製 demo，並完成分析報告，強調方法在真實機器人上的可用性。
In this summer internship, you will build a complete robot learning pipeline using LeRobot, from data collection and behavior learning to real-world deployment on a physical robot. You will benchmark policies on at least 1–2 manipulation tasks, create demos, and write a short analysis report, focusing on practical, deployable robot learning rather than just simulations.
|
其他資訊(Other Information)
|
|
主持人(PI)
逄愛君
Ai-Chun Pang
|
研究主題(Research Topic)
生成式人工智慧與多代理人系統：智慧網路管理與創新通訊技術
Generative Artificial Intelligence and Multi-Agent Systems: Intelligent Network Management and Innovative Communication Technologies
|
研究介紹(Introduction)
本實驗室專注於結合生成式人工智慧 (Generative AI) 與多代理人系統 (Multi-Agent System)，探索智慧化網路管理與通訊技術的創新應用。我們致力於開發分散式網路架構，設計多代理人間的合作與共識機制，以應對動態網路環境中的資源分配和效率挑戰。為提升系統的信任性與透明度，我們引入區塊鏈技術，並運用語意通訊 (Semantic Communication) 進一步優化代理人間的資訊共享效率。此外，針對既有地面網路覆蓋限制，我們研究低軌道衛星通訊，拓展智慧網路的應用範圍。我們的研究目標是打造一個高適應性、高效能且具韌性的次世代智慧網路管理系統，為未來的智慧城市與物聯網應用提供技術支撐。
Our research topic is integrating Generative Artificial Intelligence (Generative AI) and Multi-Agent Systems (MAS) to develop advanced solutions for intelligent network management. We tackle the challenges of complex network architectures and diverse application demands by designing distributed systems for efficient resource allocation, traffic control, and fault detection.
To ensure trust and transparency, we incorporate blockchain technology, while Semantic Communication enhances data transmission by focusing on relevant information and reducing redundancy. We also apply Generative AI to mitigate semantic interference in multi-user environments, improving communication efficiency and system performance.
Additionally, we explore Low Earth Orbit (LEO) satellite communications to extend network coverage and reliability, integrating it with terrestrial networks to create a robust, cross-domain communication framework. We aim to deliver a next-generation intelligent network system, driving innovation for smart cities, IoT, and edge computing applications.
|
其他資訊(Other Information)
|
|
主持人(PI)
曹昱
Yu Tsao
|
研究主題(Research Topic)
基於AI的生理訊號 (ECG、EEG、EMG、PPG) 分析
AI-based physiological signal analysis (ECG、EEG、EMG、PPG)
|
研究介紹(Introduction)
隨著高齡化社會來臨與智慧醫療需求快速成長，生理訊號分析在疾病預警、健康監測與精準醫療中扮演關鍵角色。然而，心電圖（ECG）、腦電圖（EEG）、肌電圖（EMG）與光體積變化訊號（PPG）等生理訊號普遍受到雜訊、個體差異與量測條件變動影響，傳統訊號處理與特徵工程方法在實務應用上仍面臨準確度、泛化性與可擴展性不足等挑戰。本計畫擬建構一套以人工智慧（AI）為核心的生理訊號分析框架，整合深度學習與先進訊號處理技術，發展具備高準確性、強泛化能力與可解釋性的多模態生理訊號分析模型。研究內容將涵蓋：（1）針對 ECG、EEG、EMG 與 PPG 訊號之自動化品質評估與雜訊抑制方法；（2）結合時序建模與多尺度特徵學習的端到端深度神經網路架構；（3）跨訊號、跨任務的共享表示學習與多任務學習策略，以提升模型在不同應用場景下的適應能力。本計畫亦將重視模型可解釋性與臨床可信度，透過注意力機制與可視化分析，協助醫療人員理解 AI 決策依據。預期成果包括高效能 AI 生理訊號分析演算法、公開可重現的研究成果，以及可應用於遠距醫療、穿戴式裝置與智慧健康監測系統之核心技術，為智慧醫療與健康照護提供具體且長期的技術支撐。
With the rapid growth of smart healthcare and an aging society, physiological signal analysis is critical for disease early warning, health monitoring, and precision medicine. However, ECG, EEG, EMG, and PPG signals are often affected by noise, inter-subject variability, and changing measurement conditions, limiting traditional analysis methods. This project aims to develop an AI-centered framework that integrates deep learning with advanced signal processing to achieve accurate, robust, and interpretable multimodal physiological signal analysis. The research focuses on automated signal quality assessment, noise suppression, end-to-end temporal and multi-scale learning architectures, and cross-signal multi-task learning. The outcomes will support reliable deployment in telemedicine, wearable devices, and intelligent health monitoring systems.
|
其他資訊(Other Information)
|
|
主持人(PI)
黃瀚萱
Hen-Hsen Huang
|
研究主題(Research Topic)
知識庫中的反事實因果關係之建立與推理
Counterfactual Causal Analysis in Knowledge Bases
|
研究介紹(Introduction)
知識圖譜由呈現事實的三元組構成，表達實體之間的關係。在典型的知識圖譜中，所有的事實都預設是可靠、真實的，但在現實上，知識圖譜很可能包含不確定的資訊，其中錯誤的事實甚至可能和其他事實互相衝突。在這個計畫中，我們預期將反事實知識引入知識圖譜，對不確定資訊進行反事實因果分析，藉以偵測與更正知識圖譜上不可靠的內容。除了可以確保知識圖譜的一致性，還可以進一步與深度學習模型的工作憶體整合，讓線上模型自動更正不實資訊。
Typical knowledge bases are composed of factual triples, representing the relations among entities with an assumption in mind that all the facts are true and reliable. In the real world, however, a knowledge base possibly contains uncertain information, and the untrue facts may be inconsistent with others mutually. In this project, our goal is to introduce a different kind of knowledge, the counterfactual knowledge, into knowledge bases to advance the causal analysis over the uncertain information. The results are expected to be not only useful for guarding the knowledge base integrity but also having the potential for misinformation correction in the machine's working memory.
|
其他資訊(Other Information)
|
|
主持人(PI)
呂俊賢
Chun-Shien Lu
|
研究主題(Research Topic)
深度學習安全與隱私
Deep Learning Security and Privacy
|
研究介紹(Introduction)
My recent research interests focus on the (Deep) Compressive Sensing and AI Security & Privacy issues. Compressive Sensing has attracted much attention due to its new paradigm of simultaneous sampling and compression. Our representative results are describe below.
Topic 1: Distributed Compressive Sensing (DCS)
Distributed compressive sensing (DCS) is a framework that considers joint sparsity within signal ensembles along with multiple measurement vectors (MMVs). However, current theoretical bounds of the probability of perfect recovery for MMVs are derived to be essentially identical to that of a single MV (SMV); this is because characteristics of the signal ensemble are ignored. In this work, our contribution is to complete the proof in that, by taking the size of signal ensembles into consideration, MMVs indeed exhibit better performance than SMV.
Sung-Hsien Hsieh, Wei-Jie, Liang, Chun-Shien Lu, and Soo-Chang Pei, ``Distributed Compressive Sensing: Performance Analysis with Diverse Signal Ensembles,’’ IEEE Trans. on Signal Processing, vol. 68, pp. 3500-3514, 2020.
Topic 2: Compressed Sensing of Large-Scale Images
Cost-efficient compressive sensing of large-scale images with quickly reconstructed high-quality results is very challenging. In this work, we present an algorithm to solve convex optimization via the tree structure sparsity pattern, which can be run in the operator to reduce computation cost and maintain good quality, especially for large-scale images.
Wei-Jie, Liang, Gang-Xuan Lin, and Chun-Shien Lu, ``Tree Structure Sparsity Pattern Guided Convex Optimization for Compressive Sensing of Large-Scale Images,’’ IEEE Trans. on Image Processing, Vol. 26, No. 2, pp. 847-859, 2017.
Topic 3: Compressed Sensing-Based Clone Identification in Sensor Network
Clone detection, aimed at detecting illegal copies with all of the credentials of legitimate sensor nodes, is of great importance for sensor networks because of the severe impact of clones on network operations. Various detection methods have been proposed, but most of them are communication-inefficient. In view of the sparse characteristic of replicated nodes, we propose a novel clone detection framework, called CSI, based on compressed sensing.
Chia-Mu Yu, Chun-Shien Lu, and Sy-Yen Kuo, “Compressed Sensing-Based Clone Identification in Sensor Network,’’ IEEE Trans. on Wireless Communications, Vol. 15, No. 4, pp. 3071-3084, 2016.
The amount of publications, pertaining to AI Security and Privacy, have been grown exponentially since 2014. This indicates that the issues of AI security and privacy has received much attention recently. Our representative results are describe below.
Topic 1: Difference-Seeking Generative Adversarial Network--Unseen Sample Generation
Unseen data, which are not samples from the distribution of training data and are difficult to collect, have exhibited importance in numerous applications, (e.g., novelty detection, semi-supervised learning, and adversarial training). In this paper, we introduce a general framework called difference-seeking generative adversarial network (DSGAN), to generate various types of unseen data.
Yi-Lin Sung, Sung-Hsien Hsieh, Soo-Chang Pei, and Chun-Shien Lu, ``Difference-Seeking Generative Adversarial Network--Unseen Sample Generation,’’ International Conference on Learning Representations (ICLR), Addis Ababa, Ethiopia, April 26-30, 2020.
Topic 2: Perceptual Differential Privacy in Images
With the growing use of camera devices, the industry has many image datasets that provide more opportunities for collaboration between the machine learning community and industry. However, the sensitive information discourage data owners from releasing these datasets. In our work, we propose perceptual indistinguishability (PI) as a formal privacy notion particularly for images. We then propose PI-Net, a privacy-preserving mechanism that achieves image obfuscation with PI guarantee.
Jia-Wei Chen, Li-Ju Chen, Chia-Mu Yu, and Chun-Shien Lu, ``Perceptual Indistinguishability-Net (PI-Net): Facial Image Obfuscation with Manipulable Semantics,’’ CVPR, June 19-25, 2021.
2. Future Plans
Following our experiences in compressive sensing, deep learning, and AI/Multimedia security & privacy, we have two future works.
Deep Sensing: Deep sensing is a new learning model in solving optimization problems for multi-tasks. To this end, we integrate deep learning and compressive sensing to solve the image inverse problems, including compressive sensing, super-resolution, inpainting, and so on. To our knowledge, we have not found any work in the literature that studied such optimization-based neural network to solve the generalized multi-task inverse problems. The challenge is, for each task, how to approximate its state-of-the-art performance.
AI Security with Robust Training: Considering the well-trained NN models should possess larger certified radii, we investigate ``Smoothed Robust Training’’ in three possible ways. First, we will design a NN model to have small Lipschitz constant, despite that it is challenging to evaluate the Lipschitz constant for each layer. Second, the training strategy is concerned in that we aim to achieve robust training by virtue of data augmentation. Finally, in view of the importance of loss function design, we plan to incorporate the optimization problem in solving the certified radius with loss function to maximize certified radius.
My recent research interests focus on the (Deep) Compressive Sensing and AI Security & Privacy issues. Compressive Sensing has attracted much attention due to its new paradigm of simultaneous sampling and compression. Our representative results are describe below.
Topic 1: Distributed Compressive Sensing (DCS)
Distributed compressive sensing (DCS) is a framework that considers joint sparsity within signal ensembles along with multiple measurement vectors (MMVs). However, current theoretical bounds of the probability of perfect recovery for MMVs are derived to be essentially identical to that of a single MV (SMV); this is because characteristics of the signal ensemble are ignored. In this work, our contribution is to complete the proof in that, by taking the size of signal ensembles into consideration, MMVs indeed exhibit better performance than SMV.
Sung-Hsien Hsieh, Wei-Jie, Liang, Chun-Shien Lu, and Soo-Chang Pei, ``Distributed Compressive Sensing: Performance Analysis with Diverse Signal Ensembles,’’ IEEE Trans. on Signal Processing, vol. 68, pp. 3500-3514, 2020.
Topic 2: Compressed Sensing of Large-Scale Images
Cost-efficient compressive sensing of large-scale images with quickly reconstructed high-quality results is very challenging. In this work, we present an algorithm to solve convex optimization via the tree structure sparsity pattern, which can be run in the operator to reduce computation cost and maintain good quality, especially for large-scale images.
Wei-Jie, Liang, Gang-Xuan Lin, and Chun-Shien Lu, ``Tree Structure Sparsity Pattern Guided Convex Optimization for Compressive Sensing of Large-Scale Images,’’ IEEE Trans. on Image Processing, Vol. 26, No. 2, pp. 847-859, 2017.
Topic 3: Compressed Sensing-Based Clone Identification in Sensor Network
Clone detection, aimed at detecting illegal copies with all of the credentials of legitimate sensor nodes, is of great importance for sensor networks because of the severe impact of clones on network operations. Various detection methods have been proposed, but most of them are communication-inefficient. In view of the sparse characteristic of replicated nodes, we propose a novel clone detection framework, called CSI, based on compressed sensing.
Chia-Mu Yu, Chun-Shien Lu, and Sy-Yen Kuo, “Compressed Sensing-Based Clone Identification in Sensor Network,’’ IEEE Trans. on Wireless Communications, Vol. 15, No. 4, pp. 3071-3084, 2016.
The amount of publications, pertaining to AI Security and Privacy, have been grown exponentially since 2014. This indicates that the issues of AI security and privacy has received much attention recently. Our representative results are describe below.
Topic 1: Difference-Seeking Generative Adversarial Network--Unseen Sample Generation
Unseen data, which are not samples from the distribution of training data and are difficult to collect, have exhibited importance in numerous applications, (e.g., novelty detection, semi-supervised learning, and adversarial training). In this paper, we introduce a general framework called difference-seeking generative adversarial network (DSGAN), to generate various types of unseen data.
Yi-Lin Sung, Sung-Hsien Hsieh, Soo-Chang Pei, and Chun-Shien Lu, ``Difference-Seeking Generative Adversarial Network--Unseen Sample Generation,’’ International Conference on Learning Representations (ICLR), Addis Ababa, Ethiopia, April 26-30, 2020.
Topic 2: Perceptual Differential Privacy in Images
With the growing use of camera devices, the industry has many image datasets that provide more opportunities for collaboration between the machine learning community and industry. However, the sensitive information discourage data owners from releasing these datasets. In our work, we propose perceptual indistinguishability (PI) as a formal privacy notion particularly for images. We then propose PI-Net, a privacy-preserving mechanism that achieves image obfuscation with PI guarantee.
Jia-Wei Chen, Li-Ju Chen, Chia-Mu Yu, and Chun-Shien Lu, ``Perceptual Indistinguishability-Net (PI-Net): Facial Image Obfuscation with Manipulable Semantics,’’ CVPR, June 19-25, 2021.
2. Future Plans
Following our experiences in compressive sensing, deep learning, and AI/Multimedia security & privacy, we have two future works.
Deep Sensing: Deep sensing is a new learning model in solving optimization problems for multi-tasks. To this end, we integrate deep learning and compressive sensing to solve the image inverse problems, including compressive sensing, super-resolution, inpainting, and so on. To our knowledge, we have not found any work in the literature that studied such optimization-based neural network to solve the generalized multi-task inverse problems. The challenge is, for each task, how to approximate its state-of-the-art performance.
AI Security with Robust Training: Considering the well-trained NN models should possess larger certified radii, we investigate ``Smoothed Robust Training’’ in three possible ways. First, we will design a NN model to have small Lipschitz constant, despite that it is challenging to evaluate the Lipschitz constant for each layer. Second, the training strategy is concerned in that we aim to achieve robust training by virtue of data augmentation. Finally, in view of the importance of loss function design, we plan to incorporate the optimization problem in solving the certified radius with loss function to maximize certified radius.
|
其他資訊(Other Information)
|
|
主持人(PI)
莊庭瑞
Tyng-Ruey Chuang
|
研究主題(Research Topic)
研究資料基礎設施與服務
Research Data Infrastructures and Services
|
研究介紹(Introduction)
研究資料寄存所實驗室 (depositar lab) 致力於研究資訊系統與工具，發展新興的研究資料基礎設施與服務。我們營運「研究資料寄存所」（研究資料的開放儲存庫 https://data.depositar.io/ ）以及「研究資料管理推進室」(RDM Hub https://rdm.depositar.io/ ) ，服務台灣和世界各地的研究人員，無論其學科領域。我們與夥伴們並進行大規模的數位保存與資料協作專案。我們使用並開發開放原始碼軟體。我們所提供的服務，所有人皆可自由使用。
我們的工作受到中央研究院（資訊科學研究所以及資訊科技創新研究中心）和國家科學與技術委員會（自然科學與永續研究發展處）的支持，部份經費來自其他單位。
實驗室位於台北南港山腳。我們的夥伴（過去與現在）包括中央研究院的地理資訊科學研究專題中心、國立台灣歷史博物館、台灣生物多樣性資訊機構 (TaiBIF)、農業部生物多樣性研究所、台灣長期社會生態核心觀測站 (LTSER Taiwan)、拾穗者文化股份有限公司、觀自然生態環境顧問有限公司以及其他單位。
過去暑期生的訓練課程與專案請參見：
https://lab.depositar.io/zh-tw/news/250630_2/
https://lab.depositar.io/zh-tw/news/240702_1/
https://lab.depositar.io/zh-tw/news/240304_1/
亦請關注本實驗室消息頁面：
https://lab.depositar.io/zh-tw/news/
The depositar lab researches and develops systems and tools for novel research data infrastructures and services. We operate the depositar ( https://data.depositar.io/ ), a public repository for research data, and curate the Research Data Management Hub (RDM Hub; https://rdm.depositar.io/ ) for researchers of all disciplines in Taiwan and worldwide. We also work with our partners on large-scale digital preservation and data collaboration projects. We use and make open source software. The services we provide are free to all to use.
Our works are supported by Taiwan's Academia Sinica (the Institute of Information Science and the Research Center for Information Technology Innovation), the National Science and Technology Council (the Department of Natural Sciences and Sustainable Development), and grants from other sources.
We are based at a hillside in Nangang, Taipei. Our partners, past and present, include the Center for GIS of Academia Sinica, National Museum of Taiwan History, Taiwan Biodiversity Information Facility (TaiBIF), Taiwan Biodiversity Research Institute, Taiwan Long-Term Social-Ecological Research Network (LTSER Taiwan), Word Gleaners Ltd., Nature Watch Ecological and Environmental Consultancy Ltd., among others.
For past internship courses and projects at the lab, please see:
https://lab.depositar.io/news/250630_2/
https://lab.depositar.io/news/240702_1/
https://lab.depositar.io/news/230711_1/
Please follows the news page of our lab:
https://lab.depositar.io/news/
|
其他資訊(Other Information)
|
|
主持人(PI)
王建堯
Chien Yao Wang
|
研究主題(Research Topic)
開放模態世界模型研究
Open-modality World Model Research
|
研究介紹(Introduction)
研究世界模型與空間智慧，支援包含但不限視覺與語言模態做為輸入。
Research world models and spatial intelligence, supporting input including but not limited to vision and language modalities.
|
其他資訊(Other Information)
|
|
主持人(PI)
廖弘源
Mark Liao
|
研究主題(Research Topic)
超高解析度3D生醫影像切割與對位研究
Ultra-high Resolution 3D Medical Image Segmentation and Registration Research
|
研究介紹(Introduction)
研究超高解析度3D影像分析技術，主要以層光影像的分割與對位技術為主。
The research focuses on ultra-high resolution 3D image analysis technology, mainly on the segmentation and alignment techniques of light sheet images.
|
其他資訊(Other Information)
|
|
主持人(PI)
葉彌妍
Mi-Yen Yeh
|
研究主題(Research Topic)
深度學習與大語言模型於人工智慧應用
Deep Learning and Large Language models for AI applications
|
研究介紹(Introduction)
運用深度學習與大語言模型於各式人工智慧應用，例如大語言模型與演算法設計(少樣本學習、模型融合與效能優化等)、大語言模型代理人應用及能力評測資料搜集與生成、大語言模型與深度學習模型應用於智慧交易、知識問答工作等。
Leveraging deep learning and large language models (LLMs) across a wide range of AI applications, including LLM-based algorithm and system design (e.g., few-shot learning, model merging, and performance optimization), LLM agent applications and capability evaluation through data collection and generation, as well as applying LLMs and deep learning models to intelligent trading and knowledge-question answering tasks.
|
其他資訊(Other Information)
|
|
主持人(PI)
鄭湘筠
Hsiang-Yun Cheng
|
研究主題(Research Topic)
記憶體內深度學習與大數據分析之軟硬體協同設計
Software-hardware co-design for memory-centric deep learning and data analytics
|
研究介紹(Introduction)
近年來，大數據分析（如深度學習、圖論分析與基因序列分析等）迅速發展，這些應用在運算中仰賴高效能的巨量數據存取，現有主流運算系統卻往往無法滿足此需求，促使電腦系統設計必須重新思考。以記憶體為中心的運算架構因而成為極具潛力的設計方向，透過在記憶體內或其周邊直接執行運算，以降低資料搬移所造成的效能與耗能瓶頸。多種新興記憶體技術（如 ReRAM、PCM、MRAM、FeFET 及 NAND/NOR Flash）已展現結合儲存與運算的能力，並可支援矩陣向量乘法、邏輯運算與向量相似度搜尋等應用，產業界亦積極推動近記憶體運算技術（如 Samsung 的 HBM-PIM 和 AxDIMM、SK Hynix 的 AiM、UPMEM 的 PIM 等）。然而，受限於硬體技術尚未成熟，運算模式與傳統架構差異大，以及各類大數據分析演算法在運算與存取特性上的高度異質性，在系統設計上仍面臨眾多挑戰，有待進一步克服。
本實習計畫的目標為針對大數據分析的多元應用情境，系統性探討不同層面的設計挑戰，涵蓋電路與元件階層、計算架構階層以及演算法階層，並透過軟硬體協同設計，充分發掘記憶體內及近記憶體運算的優勢，開發高效能、低耗電的新世代運算系統。此外，由於此類系統具備低功耗特性，有助於環境永續發展與減少碳排放，我們也歡迎實習生進一步探討此研究方向。
實習生可選擇參與下列研究主題，或其他相關研究議題。
1. 透過軟硬體協同設計，以記憶體內及存儲內運算，實現高效能低耗電之深度學習，包括生成式AI、大型語言模型、推薦系統等。
2. 針對具有不規則數據存取及複合式運算行為之大數據分析應用情境，如資料樣式探勘、基因序列比對等，設計異質性記憶體為中心運算系統，並優化資料配置與運算排程。
3. 探討以chiplet方式異直整合運算單元(CPU, GPU, TPU, etc)與具備運算能力之記憶體或存儲單元(SRAM, HBM, Flash, etc)，在各個系統層級面臨的挑戰，並以軟硬體協同設計，開發可能的解決方式。
4. 探討以記憶體為中心運算及Chiplet異直晶片整合方式對碳排放量之影響，並開發對環境永續發展友善之運算系統。
Data-intensive applications such as deep learning, graph analytics, and genome analysis increasingly demand efficient access to massive data volumes, exposing the limitations of traditional processor-centric computing systems. Memory-centric computing offers a promising alternative by performing computations within or near memory to reduce data-movement overhead. Enabled by emerging memory technologies such as ReRAM, PCM, MRAM, FeFET, and NAND/NOR Flash, memory-centric systems can support parallel operations including matrix–vector multiplication, bitwise logic, and vector similarity search. Industry has also actively developed near-memory computing solutions by integrating simple compute units into 3D-stacked memory or DRAM modules, such as Samsung’s HBM-PIM and AxDIMM, SK Hynix’s AiM, and UPMEM’s PIM. However, realizing these systems in practice remains challenging due to hardware constraints and the diverse computational and data-access characteristics of applications.
Our goal is to investigate design challenges across multiple system layers, including the device/circuit, architecture, and algorithm levels, and to develop cross-layer solutions that fully exploit the potential of in-memory and near-memory computing. Moreover, given the low energy consumption of memory-centric computing and its strong potential to improve sustainability and reduce carbon emissions, we also encourage summer interns to participate in this exciting research direction.
Candidate topics include, but are not limited to, the following:
1. Energy-efficient deep learning enabled by in-memory and in-storage computing, focusing on the design of memory-centric architectures tailored for applications such as generative AI, large language models (LLMs), recommendation systems, and graph neural networks.
2. Design heterogeneous memory-centric computing systems for applications with irregular data access patterns and complex computational behaviors (e.g., graph pattern mining and genomic sequence analysis), with an emphasis on optimized data mapping and computation scheduling.
3. Explore the challenges across different system levels in heterogeneously integrating compute units (e.g., CPU, GPU, TPU) with compute-capable memory or storage units (e.g., SRAM, HBM, Flash) using a chiplet-based approach, and develop potential solutions through hardware–software co-design.
4. Analyze the potential benefits and challenges of memory-centric computing and chiplet-based heterogeneous integration for improving sustainability and reducing carbon emissions, and develop memory-centric systems that are aligned with these sustainability goals.
|
其他資訊(Other Information)
|
|
主持人(PI)
王建民
Chien-Min Wang
|
研究主題(Research Topic)
機器學習技術與應用
Machine Learning Technologies and Applications
|
研究介紹(Introduction)
(1) 多代理人強化學習：針對多代理人強化學習與合作決策之行為，於多種複雜且具高度動態特性的模擬環境中進行系統性的探討。隨著以強化學習為基礎之多代理人方法迅速發展，其在多自主體協同、集體智慧與自我調適能力方面已展現出極大潛力。然而這類方法在長期訓練下的行為演變、對環境變化的反應模式、以及在不同任務複雜度下的穩定性與可依賴性，仍缺乏全面且深入的研究。本研究將針對多代理人演算法在長期訓練過程中的協作行為、策略形成、適應性調整與決策穩健性進行分析，特別著重於隨著環境目標、任務結構與條件變化時，其學習效果與系統反應之差異。透過多面向的觀察與比較，本計畫期望建構一套更完整的多代理人行為理解模型，呈現其在跨場景條件下的學習模式與協作機制。
(2) 時間序列預測：過去研究的成果中，我們研究將機器學習應用於時間序列預測的方法，並和傳統的時間序列預測方法進行比較，獲得了不錯的成果。然而我們也發現了採用單一預測方法的缺陷，因為沒有任何一種時間序列預測方法能夠在所有的測試都勝過其他對手。因此目前正進行的研究將探索自動化特徵選取技術和元學習(Meta-Learning)技術，於多種時間序列預測方法中找出最佳的選擇，以進一步提升時間序列預測的準度以及應對不同特性之資料集的預測能力。
(1) Multi-agent Reinforcement Learning：This project undertakes a systematic study of cooperative multi-agent reinforcement learning (MARL) and decision-making processes within a range of complex and dynamically evolving simulation environments. Modern multi-agent learning techniques, including those derived from reinforcement learning (RL), have shown significant potential for enabling coordinated behavior among multiple autonomous entities. However, their long-term behavior, adaptability, and stability under changing task structures remain insufficiently understood. The proposed research aims to address this gap by examining how different multi-agent learning approaches develop coordinated strategies over extended training periods, how they react to shifts in environmental conditions, and how their performance evolves under varying levels of complexity. By placing emphasis on dynamic transitions, long-horizon learning characteristics, and system-level behavior patterns, the project seeks to generate a comprehensive picture of cooperative learning dynamics across diverse scenarios.
(2) Time Series Forecasting：In the results of past research, we studied the methods of applying machine learning to time series forecasting and compared them with traditional time series forecasting methods, achieving good results. However, we also identified the limitations of using a single forecasting method, as no single time series forecasting method can outperform all others in every test. Therefore, the ongoing research will explore automated feature selection techniques and meta-learning techniques to identify the best choices among various time series forecasting methods, in order to further enhance the accuracy of time series forecasting and improve the predictive capabilities for datasets with different characteristics.
|
其他資訊(Other Information)
|
|
主持人(PI)
蔡孟宗
Meng-Tsung Tsai
|
研究主題(Research Topic)
串流式圖論演算法
Graph Streaming Algorithms
|
研究介紹(Introduction)
我的研究興趣在探討如何使用 O(n) 的記憶體空間處理各式的圖論計算問題，這裡 n 是指輸入圖的節點個數。
我們假設圖的邊是按照某個最糟的順序一條一條給演算法，而且只給一次。一張 n 個節點的圖，最多會有 Ω(n^2) 條邊，因為限制只能使用 O(n) 的記憶體空間，勢必得強迫演算法 "忘記" 大部分曾經讀進來的邊。在這個前提下，如何設計演算法完成各式的圖論計算問題？
在這嚴格的限制下，或許心裡的第一問題是："是否大部分的圖論問題都不能在使用 O(n) 記憶體的狀況下完成計算？" 目前的研究文獻已經證實，許多圖論計算問題可以，但也有許多圖論計算問題，保證無法在這限制下計算出來。後者的情況，常常能找到方法在使用少量的空間下，找到 (1) 不錯的近似解、(2) 具有隨機成分的最佳解 (在很高的成功機率下)、或 (3) 具有隨機成分的不錯的近似解 (在很高的成功機率下)。這邊的機率與輸入的圖無關，只和演算法使用的隨機成分有關。
近期實驗室的研究成果有：
1. 存在一般圖上的 NP-complete 圖論計算問題，可以使用 O(n) 空間回答！
2. 對於將輸入圖拆分成盡可能少的無環子圖這個圖論計算問題，任何演算法都需要 Ω(n^2) 的記憶體空間才能找到最佳的拆分法！但存在演算法，只要 O(n) 的記憶體空間就能找到近似於最佳解的拆分法。
3. 如果演算法可以在固定的邊序列上重複讀取多次，例如 O(k) 次 for some k > 0，那麼它可以用 O(𝛼n^{1+1/k}) 空間完成拓樸排序與強連通塊分解其中 𝛼 為輸入圖的最大獨立數。
去年暑期實習生的研究成果有：
1. 給定一般圖 G，判斷 G 是否有個生成樹滿足 t-spanner 的性質，不管 t 是哪個大於 1 的整數，如果不使用隨機、而且只看圖一次，任何演算法都需要 Ω(n^2) 的空間。
2. 給定 R^3 空間中兩群點，判斷兩群點是否同構可以在看過點群常數次後，用 O(n^r) 空間回答，其中 n 為點群中的點數、r 為小於 1 的某常數。
在這個專題，我們預期可以學習到如何使用數學工具回答："在侷限的記憶體空間下，有哪些圖論計算問題可以被解決？有哪些圖論計算問題保證無法被解決？以及你喜歡的圖論計論問題是屬於哪一類？"
We are interested in whether graph problems on n-node inputs can be solved in O(n) space.
We assume that the edges of the input graph are given to the algorithm one by one, in an arbitrary order, and only once. Note that an n-node graph may have Ω(n^2) edges. If an algorithm uses O(n) space, then it has to ``forget'' much information of the input. Given the restriction, can we design algorithms to solve graph problems?
One may wonder whether there are many problems that can be solved using little space. It has been shown in the literature that: dozens of graph problems can be solved using little space, while dozens of graph problems cannot. In the latter case, the community usually can come up with a solution that approximates the best possible to within some factor, a solution that matches an optimal one with high probability, or a solution that approximates the best possible to within some factor with high probability. The probabilities here depend only on the randomness used in algorithms, and do not depend on the input graph.
The recent results obtained by our lab include:
1. There exists some NP-complete graph problem on general graphs that can be computed using O(n) space!
2. For any streaming algorithm, decomposing a graph into the least number of acyclic subgraphs requires Ω(n^2) space. However, this problem can be well approximated using O(n) space.
3. If the algorithm is allowed to scan the edges of the input graph multiple times in a fixed order, say O(k) times, then it can compute a topological ordering and decompose the nodes into strongly connected components using O(𝛼n^{1+1/k}) space, where 𝛼 is the independence number of the input graph.
The results obtained by the summer interns last year include:
1. For each integer t >= 2, any deterministic single-pass streaming algorithm for finding a tree t-spanner for a given n-node undirected simple graph requires Ω(n^2) space.
2. Given two sets of n points in R^3, there exists an O(1)-pass streaming algorithm that test the congruence of these two sets using O(n^r) space for some constant r < 1.
In this independent study, we expect to learn how to apply mathematical methods to answer the questions: whether a graph problem can be solved using little space, and which category your favorite graph problem belongs to?
|
其他資訊(Other Information)
|
|
主持人(PI)
馬偉雲
Wei-Yun Ma
|
研究主題(Research Topic)
LLM 的持續學習與記憶力與多模態LLM架構
LLM Continual Learning, Memory Capabilities, and Multimodal Architectures
|
研究介紹(Introduction)
在今年暑假，我們開放數個名額給實習生，一起參與以下這兩個有趣又充滿挑戰的主題。
* 主題1: LLM 的持續學習與記憶力
大型語言模型 (LLMs) 的致命傷在於知識截止 (Knowledge Cutoff) 與災難性遺忘 (Catastrophic Forgetting)。當前 RAG 技術雖能外掛知識，但無法讓模型真正「內化」新知。在此主題中，我們同步推進兩大前沿技術路徑：
1. 架構層面的記憶閘門 (Architectural Approach)： 我們獨創 LoRA Output Scaling Gate 機制，在模型內部植入動態閘門，讓模型能自主判斷並切換「新知識 LoRA」與「舊知識 Backbone」的權重貢獻。透過讓 LoRA 模組專精於新資訊 (Learn more) 並由主模型維持既有能力 (Forget less)，我們在 GSM8K 等評測上已證實能顯著降低遺忘率 。
2. 參數層面的模型融合 (Model Merging Approach)： 我們跟進開源社群與學術界最活躍的 Continual LoRA 方向，實踐 "Merge before Forget" 策略。研究如何利用 TIES-Merging 或 DARE 等演算法，將針對新任務訓練的 LoRA 權重，以數學方式精確合併回主模型或整合為「超級 Adapter」。同時，我們引入 PS-LoRA (Parameter Stability LoRA) 概念，在訓練階段即施加 正交性約束 (Orthogonal Constraints)，確保新的參數更新方向與既有知識空間互不衝突，從數學本質上解決知識干擾問題。
* 主題2: 全新的語音與文字的多模態LLM架構
要打造一個結合語音辨識和語言理解系統，傳統的流水線 (pipeline) 方法是先利用 ASR (自動語音辨識) 將輸入語音逐字轉換為文字，再通過 NLU (自然語言理解) 抽取語義。然而，這種分段處理方式會導致錯誤傳遞 (error propagation) 問題。舉例而言，當老年人與照護 AI 或虛擬醫護人員互動時，可能因語音斷斷續續或發音不清，且台語與國語交替，造成句子獨立來看時語意模糊。真人通常能憑藉世界知識和常識，結合上下文推測對方的意思，但流水線方式卻面臨巨大挑戰。缺乏世界知識、常識與上下文的情境下，對於方言、口音甚至不連貫的語音訊號，傳統 ASR 的辨識表現會顯著下降，哪怕只是幾個字的錯誤，也會嚴重影響後續NLU的效果。此外，流水線方法的另一大先天限制是，當 ASR 將語音訊號轉為文字後，NLU 無法再利用語音中的其他訊息，如情緒、語調、口氣與停頓，這些可能影響語意理解的信號。例如，“你還真行啊”在真誠與挖苦的語氣下，語意截然不同。唯有多模態模型通過綜合判讀，才能精準捕捉這些細微差異，做出合適的回應。
我們目前正在開發一套全新的語音與文字的多模態LLM架構，提出能達到realtime的新方法，已經有初步成果。
(本實驗室的最新的研究工作之一 : 動態調整LLM每筆訓練資料的「重要性」以提升整體訓練效果，榮登國際頂會ICLR 2025, 並獲選為Spotlight paper, 為台灣在 LLM 領域首次獲得 ICLR Spotlight 殊榮，展現研究的原創性與國際影響力。)
This summer, we are opening several internship positions to work with us on two exciting and challenging research topics.
Topic 1: Continual Learning and Memory in LLMs
The critical limitations of Large Language Models (LLMs) lie in Knowledge Cutoff and Catastrophic Forgetting. While current RAG (Retrieval-Augmented Generation) technologies allow for external knowledge retrieval, they fail to enable models to truly "internalize" new information. In this topic, we are simultaneously advancing two frontier technical approaches:
1. Architectural Approach (Memory Gating): We have developed a novel LoRA Output Scaling Gate mechanism. By embedding dynamic gates within the model, it autonomously evaluates and adjusts the weight contributions of the "New Knowledge LoRA" versus the "Old Knowledge Backbone." This allows the LoRA module to specialize in new information (Learn more) while the main model maintains existing capabilities (Forget less). Our benchmarks on datasets like GSM8K have confirmed a significant reduction in forgetting rates.
2. Parameter-Level Approach (Model Merging): Aligning with the most active Continual LoRA research in the open-source and academic communities, we are implementing a "Merge before Forget" strategy. We are researching how to utilize algorithms like TIES-Merging or DARE to mathematically merge LoRA weights trained on new tasks back into the main model or integrate them into a "Super Adapter." Additionally, we are introducing the concept of PS-LoRA (Parameter Stability LoRA). By applying Orthogonal Constraints during training, we ensure that new parameter updates do not conflict with the existing knowledge space, mathematically solving the issue of knowledge interference.
Topic 2: A Novel Speech-Text Multimodal LLM Architecture
Traditionally, building a system that combines speech recognition and language understanding relies on a pipeline approach: using ASR (Automatic Speech Recognition) to transcribe speech into text, followed by NLU (Natural Language Understanding) to extract meaning. However, this segmented process suffers from error propagation.
For instance, when elderly users interact with caregiving AI or virtual medical assistants, their speech may be intermittent, unclear, or alternate between Taiwanese and Mandarin (code-switching). While humans can infer meaning using world knowledge, common sense, and context, pipeline methods struggle significantly. Without access to broader context, traditional ASR performance drops sharply with dialects, accents, or disjointed speech signals. Even a few transcription errors can severely impact subsequent NLU performance.
Furthermore, an inherent limitation of the pipeline approach is the loss of paralinguistic information. Once ASR converts speech to text, the NLU loses access to cues like emotion, intonation, tone, and pauses—signals that are crucial for semantic understanding. For example, the phrase "You're really something" has vastly different meanings when spoken sincerely versus sarcastically. Only a multimodal model that synthesizes all these signals can accurately capture these nuances and respond appropriately.
We are currently developing a novel Speech-Text Multimodal LLM architecture, proposing a new method capable of real-time performance, and have already achieved promising preliminary results.
(Lab Highlight: One of our latest research works—dynamically adjusting the "importance" of each LLM training data point to improve overall training efficiency—has been accepted to the top-tier international conference ICLR 2025 and selected as a Spotlight paper. This marks the first time a research team from Taiwan has received an ICLR Spotlight in the LLM field, demonstrating the originality and international impact of our research.)
|
其他資訊(Other Information)
|
|
主持人(PI)
陳郁方
Yu-Fang Chen
|
研究主題(Research Topic)
形式化方法在量子軟體上的相關研究
Topics on Formal Methods in Quantum Software
|
研究介紹(Introduction)
我們的研究室專注於形式化方法。形式化方法是一類嚴謹的理論與技術，用於確保軟體與硬體系統的正確性與可靠性。目前，我們的研究重點放在量子軟體相關議題，主要包括以下方向。
量子程式驗證
發展並推進適用於量子計算的形式化驗證技術，處理量子疊加、糾纏與測量所帶來的理論與實務挑戰。
量子電路與量子程式的經典模擬
研究並發展用於量子電路模擬的形式化工具與表示方法，例如決策圖、自動機與相關的符號化技術，以支援可擴展的分析與驗證。
量子電路的編譯與合成
研究量子電路與量子程式在不同抽象層級之間的轉換與最佳化，包含從高階量子程式到低階量子電路的編譯流程，以及在保持語意正確性的前提下進行電路合成與優化。
實習內容
作為我們的實習生，您將參與下列一項或多項活動。
閱讀並分析相關學術論文，理解形式化方法與量子計算交會領域的最新研究進展。
參與小型研究專題，實際動手解決具體的形式化驗證或分析問題。
參與研究室討論、研討會或學術交流活動，分享研究發現，並與其他研究者進行互動。
Our laboratory focuses on formal methods. Formal methods are a class of rigorous theories and techniques used to ensure the correctness and reliability of software and hardware systems. Our current research emphasis is on quantum software, with a primary focus on the following directions.
*Quantum program verification
We develop and advance formal verification techniques tailored to quantum computing, addressing the theoretical and practical challenges arising from quantum superposition, entanglement, and measurement.
*Classical simulation of quantum circuits and quantum programs We study and develop formal tools and representation methods for simulating quantum circuits, such as decision diagrams, automata, and related symbolic techniques, to support scalable analysis and verification.
*Compilation and synthesis of quantum circuits We investigate the translation and optimization of quantum circuits and quantum programs across different abstraction levels, including compilation pipelines from high level quantum programs to low level quantum circuits, as well as circuit synthesis and optimization while preserving semantic correctness.
Internship Content:
As an intern in our lab, you will be involved in one or more of the following activities:
-Read and analyze relevant academic papers to grasp the latest developments in formal verification.
-Participate in small-scale research projects, tackling real-world formal verification challenges.
-Attend workshops and academic exchanges, sharing your findings and interacting with fellow researchers.
|
其他資訊(Other Information)
|
|
主持人(PI)
黃彥男
Yen-Nun Huang
|
研究主題(Research Topic)
(1) 資安防禦與韌性, (2) 智慧綠能機房與雲端計算, (3) AI 隱私與安全
(1) Intrusion Detection and Defense Resilience, (2) Smart Green Data Center and Cloud Computing, (3) AI 隱私與安全 (AI Privacy and Security)
|
研究介紹(Introduction)
【主題一：入侵偵測與防禦韌性研究：系統研發、對抗例生成與自動化規則】
本實習專注於提升網路安全防禦的韌性與自動化能力，實習生將參與入侵偵測系統（IDS）的研發與測試。重點工作包含研發對抗例生成技術（Adversarial Example Generation）以測試並強化模型強健性，利用 AI 技術自動生成入侵偵測規則，並運用人工智慧框架分析系統稽核紀錄（Log）與網路封包（Packet）。建議具備能力：熟悉 Python 程式語言，具備計算機網路（Network Protocols）與資訊安全基礎，若有深度學習框架（PyTorch/TensorFlow）使用經驗者尤佳。
【主題二：智慧綠能機房之數位分身、強化學習控制與雲端資源管理】
本計畫目標為提升機房自然冷源利用並降低整體能耗，實習內容強調「落地整合」。研究方向包含：(1) 數位分身：使用 Joint-PCNN 進行環境溫濕度預測；(2) 強化學習（RL）：優化空調與混風控制策略；(3) Kubernetes 資源管理：能源感知排程與效能優化。實習生將協助將感測數據、模型預測與控制決策串聯成端到端系統。
建議具備能力：熟悉 Python (Numpy/Pandas) 與資料分析。根據有興趣的分項，具備 PyTorch/TensorFlow (建模)、RL 基礎概念 (控制)、或 Docker/Kubernetes (系統) 相關背景者優先錄取。
【主題三：聯邦式學習環境下之生成式 AI 隱私攻擊防禦研究】
本實習將探討在聯邦式學習架構中，針對生成式 AI 模型（如 LLMs）的隱私洩漏風險進行分析。實習生將協助模擬模型反轉攻擊（Model Inversion Attacks）場景，並評估差分隱私（Differential Privacy）等防禦機制的有效性。
建議具備能力：適合對 AI 安全與隱私計算有興趣的同學。需熟悉 Python 與 PyTorch/TensorFlow 框架，具備機器學習基礎知識，有閱讀學術論文經驗者尤佳。
【Topic 1: Intrusion Detection and Defense Resilience: System Development, Adversarial Attacks, and Automated Rule Generation】
This internship focuses on enhancing the resilience and automation of cybersecurity defenses, where interns will participate in the development and testing of Intrusion Detection Systems (IDS). Key responsibilities include developing adversarial example generation techniques to test and improve model robustness, utilizing AI to automatically generate intrusion detection rules, and applying AI frameworks to analyze system audit logs and network packets. Requirements: Proficiency in Python; fundamental knowledge of computer networks and information security. Experience with deep learning frameworks (PyTorch/TensorFlow) is a plus.
【Topic 2: Smart Green Data Center: Digital Twin, RL Control, and Cloud Resource Management】
This project aims to optimize energy efficiency in data centers through "practical integration." Research tracks include: (1) Digital Twin: Using Joint-PCNN for environmental prediction; (2) Reinforcement Learning (RL): Optimizing HVAC and cooling control strategies; (3) Kubernetes Resource Management: Energy-aware scheduling and performance optimization. Interns will help integrate sensing data, model predictions, and control decisions into an end-to-end system.
Requirements: Proficiency in Python (Numpy/Pandas) and data analysis. Depending on the chosen track, background in PyTorch/TensorFlow (Modeling), RL concepts (Control), or Docker/Kubernetes (System) is preferred.
【Topic 3: Defense Strategies against Privacy Attacks on Generative AI in Federated Learning】
This internship investigates privacy risks associated with Generative AI models within Federated Learning architectures. Interns will assist in simulating Model Inversion Attacks and evaluating the effectiveness of defense mechanisms such as Differential Privacy.
Requirements: Ideal for students interested in AI security and privacy-preserving computing. Proficiency in Python and PyTorch/TensorFlow; basic knowledge of Machine Learning. Experience in reading academic papers is a plus.
|
其他資訊(Other Information)
|
|
主持人(PI)
王志宇
Chih-Yu Wang
|
研究主題(Research Topic)
無線網路/邊緣智慧/量子網路
Wireless Network / Edge Intelligence / Quantum Network
|
研究介紹(Introduction)
從事無線網路與邊緣智慧(含IoT，Edge Intelligence)、量子網路(Quantum Network)等相關研究，以學術論文發表與prototype系統實作為目標。
在實習開始前會先進行職前訓練，讓實習生先備妥背景知識和相關技能，實習中會提供充份資源與討論，以期待實習期間能有完整的研究體驗。如學生在實習期間有實質研究成果，本實驗室會提供專任/兼任研究助理職位以讓學生持續進行並完成研究。
We are seeking interns who are interested in tackling latest topics in wireless networks, edge intelligence, and quantum networks. Our goal is to establish academic publications and prototyping.
Resource for proper pre-training will be provided for those who are willing to contribute to the latest challenges in these research areas. Students who make promising progresses during the internship can receive follow-up RA offers if they wish to continue their research.
|
其他資訊(Other Information)
|
|
主持人(PI)
王有德
Yu Te Wang
|
研究主題(Research Topic)
(1)腦波記事本 (2)腦紋- 使用腦波來登入穿戴式裝置
(1)BrainNoter (2)BrainPrint- person authentication for wearable devices (XR)
|
研究介紹(Introduction)
歡迎來到腦機介面（BCI）世界！今年夏天，我們將尋找對BCI與混合實境（XR）感興趣的同學。我們將使用您的生理信號（如大腦、肌肉、或眼球運動軌跡）透過 XR 與週邊設備進行通信或互動。例如，您可以透過XR頭戴裝置，使用您的大腦信號來輸入和發送訊息！讓我們一同期待今年夏天會有什麼驚奇！
專案敘述：
一、腦波記事本
本計畫旨在：（1）運用 3D 列印技術設計並製作智慧眼鏡原型；（2）開發並調整最先進的解碼演算法，以詮釋即時的神經活動。最終目標是提供使用者一副智慧眼鏡，使其能在資訊較難理解的情境下（例如課堂或演講期間）快速進行影像錄製。
二、腦紋- 使用腦波來登入穿戴式裝置
儘管指紋提供了一種可靠的個人身份驗證方法，近期研究顯示這種生物識別技術在未來可能不再安全。今年夏天，我們將進一步探索這一項目，通過分析生物數據（如腦電波、心電圖、眼電圖等）來開發一個穩健的深度學習模型，用於現有數據集的個人身份驗證。
您的職責：
(1)招募受試者並進行實驗以收集生物數據。
(2)與團隊成員合作，使用人工智慧（AI）、機器學習（ML）或深度學習（DL）等工具分析收集到的數據。
(3)彙整結果並發表論文。
你的技能組合：
1. 3D 印表機（FDM 或 LCD）
2. 程式設計（Python、深度學習、機器學習）
3. 資料分析
Recruitment:
Welcome to the Brain-Computer Interface (BCI) world! This summer, we are looking for students who are interested in BCI-enabled mixed reality (XR) applications. We will work together on projects that use your bio-signal (such as brain activity, muscle activity, or eyes movement) to communicate or interact with peripheral devices via XR. For instance, you might wear a XR headset to type and send messages using your brain signal!! Amazing right? Let’s see what we have this summer.
Potential project description:
1) BrainNoter
This project aims to: (1) design and fabricate a smart glasses prototype using 3D printing technology, and (2) develop and adapt state-of-the-art decoding algorithms to interpret real-time neural activity. The ultimate objective is to provide users with smart glasses that allow rapid video recording in situations where information is difficult to comprehend, for example during lectures or presentations.
2) BrainPrint- person authentication for wearable devices (XR).
Although fingerprints provide a reliable method for personal authentication, recent studies suggest that this biometric may not remain secure in the near future. This summer, we will further explore this initiative by analyzing bio-data (such as EEG, EKG, EOG, etc.) and aim to develop a robust deep learning model for personal authentication using the existing dataset.
Your responsibility:
1) Recruit human subjects and conduct experiments to collect bio-data.
2) Work with team members to analyze the collected data using AI, ML, or DL models/tools.
3) Compile the results and publish a conference paper.
Your skill sets:
3D printer (either FDM or LCD), Programming (Python, DL, ML), Data analysis
|
其他資訊(Other Information)
|
|
主持人(PI)
王協源
Shie-Yuan Wang
|
研究主題(Research Topic)
網路數位孿生與人工智慧
Network Digital Twins with Artificial Intelligence
|
研究介紹(Introduction)
網路數位孿生 (Network Digital Twin; NDT) 為一個實體網路 (physical network)的數位分身，此實體網路中的所有設備 (例如主機、伺服器、交換機、路由器等)、通訊線路 (例如電纜、微波無線傳輸通道、光纖等) 皆一對一對應到此網路拓樸(network topology) 中的點 (node) 和線 (link)。
比起過去網路管理人員直接對實體網路改變其拓樸、運作的參數值、所使用的通訊協定、或所採用的政策（policy）所伴隨而來讓實體網路無法正常運作的巨大風險，在NDT上網管人員可以使用電腦模擬技術和人工智慧技術來研究在一些"萬一 ..."的假設強況下，採用哪種網路拓樸、參數值、通訊協定、和政策（policy）可以讓實體網路達到最佳運行效能或在多個線路故障情況下更有強韌性，等確定最好的方案已經在NDT中找到後，網管人員再將之運用到對應的實體網路，如此可以大幅減少使用新方案法卻破壞真實網路正常運作的風險。
在此計畫中，我們將設計與實作與網路數位孿生和人工智慧相關的系統並實際將此系統運用於真實世界中，應用的領域將包含 AI 資料中心網路、智慧物聯網、機器人內部通訊網路、多機器人之間通訊網路等。我們的網路數位孿生系統已經開發一年多，且已經向全世界開源，其網站在 www.ndtwin.org，有意申請加入此計畫者可先上網瞭解。
Network Digital Twin (NDT) is a digital replica of a physical network, wherein all devices in the physical network (e.g., hosts, servers, switches, routers) and communication links (e.g., cables, microwave wireless transmission channels, optical fibers) are mapped one-to-one onto nodes and links in the network topology.
Compared to the traditional method where network administrators directly modify the topology, operational parameters, communication protocols, or policies of a physical network—often risking major disruptions to its operation—NDT offers a safer alternative. Using computer simulation and artificial intelligence technologies, network administrators can study "what if..." scenarios. This enables them to determine the optimal network topology, parameter configurations, communication protocols, and policies that ensure the best performance or enhance the resilience of the physical network under conditions such as multiple link failures. Once the best solution is identified within the NDT, it can then be implemented on the corresponding physical network, significantly reducing the risk of causing disruptions during the transition to the new solution.
In this project, we will design and implement systems related to network digital twins and artificial intelligence, with the goals of applying these systems in real-world scenarios such as AI datacenter networks, Internet of Things (IoT) networks, Artificial Intelligence of Things (AIoT), In-Robot and Multi-Robot Communication Networks, etc. Our NDT system has been designed and implemented for more than one year and we have released it to the community as an open source project. Its web site is at www.ndtwin.org. Applicants can visit its web site to get more information about it.
|
其他資訊(Other Information)
|