| 主持人(PI) | 研究主題(Research Topic) | 研究介紹(Introduction) | 其他資訊(Other Information) |
|---|---|---|---|
| 徐讚昇 Tsan-sheng Hsu | 資料密集運算之分析和實做 Data intensive computing: analysis and implementation | 大型對局殘局資料庫之分析和實做 We will look into some data analysis and implementation issues for the endgame databases of some stochastic classical board games. |
PI個人首頁(PI's Information) : http://www.iis.sinica.edu.tw/pages/tshsu/ 實驗室網址(Research Information) : https://chess.iis.sinica.edu.tw/lab/ https:// Email : carol@iis.sinica.edu.tw |
| 鐘楷閔 Kai-Min Chung | 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 | 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 on the questions, and discuss it with the PI in research meetings. Students interested in theoretical computer science, particularly on the abovementioned topics, are encouraged to apply and elaborate on their interests in TCS in their 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 on the questions, and discuss it with the PI in research meetings. Students interested in theoretical computer science, particularly on the abovementioned topics, are encouraged to apply and elaborate on their interests in TCS in their application. |
PI個人首頁(PI's Information) : http://www.iis.sinica.edu.tw/pages/kmchung/ 實驗室網址(Research Information) : http://www.iis.sinica.edu.tw/~kmchung/ https:// Email : kmchung@iis.sinica.edu.tw |
| 古倫維 Lun-Wei Ku | (1) 多模態問題生成 (看圖對話) (2) 假新聞免疫 (3) 運動科技-智慧教練 (1) Multimodal Question Generation (2) Fake News Immune (3) SportTech - AI Coach | 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) 在多模態問題生成專案中，我們注重在圖像(照片、影片)問題生成，以人類見到圖像時腦中建構的世界為概念，接著產生自然可開啟對話的問題。相關技術可應用於社群網站提高人氣，或是關懷病人與老人。 (2) 在假新聞免疫研究中，我們著重於研究甚麼樣的新聞內容與呈現形式，讀者會傾向於相信或不相信，我們將進行內容理解，網路模擬及使用者端的研究。 (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。 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, and generate dialogs. Machine learning and deep learning technologies for NLP will be touched. Interns can select the topic/team they wish to join. (1) In multimodal question generation project, we are focusing on the concept that human will construct the perception of a seen image sequence. Then we aim at generating a natural question to provoke the following pleasant conversations. This research is a topic extended from our series research on VIST (visual storytelling). The developed model can be used in social media platform or conversations with people who need mental care. (2) In fake news immune project, we focus on studying why and how readers trust fake news. We will explore approaches which mitigate the impact of fake news. Moreover, the goal is to provide a mechanism which can dynamically adjust the reading environment to increase the immunity to fake news. (3) In the sport technology project, we want to try applying few-shot or zero-shot learning on gestures/elements of the specific sport. Then we aim at generating automatic coaching instructions based on the videos of gestures/elements. (4) Interns can also choose to develop demo applications for the existing technologies in our lab. The research topics include but not limited to the above. After the internship, students with good performance can continue to work with the laboratory to research and publish papers. |
PI個人首頁(PI's Information) : http://www.iis.sinica.edu.tw/pages/lwku/ 實驗室網址(Research Information) : http://academiasinicanlplab.github.io/ https:// Email : lwku@iis.sinica.edu.tw |
| 黃瀚萱 Hen-Hsen Huang | 知識庫中的反事實因果關係之建立與推理 Counterfactual Causal Analysis in Knowledge Bases | 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. |
PI個人首頁(PI's Information) : http://www.iis.sinica.edu.tw/pages/hhhuang/ 實驗室網址(Research Information) : http://www.iis.sinica.edu.tw/pages/hhhuang/ https:// Email : hhhuang@iis.sinica.edu.tw |
| 楊得年 De-Nian Yang | 元宇宙之社群多媒體網路與深度學習 Multimedia Social Networks and Deep Learning in the Metaverse | (一)社群資料探勘、機器學習與演算法設計： • 基於虛擬、擴增實境(VR/AR)或元宇宙(metaverse)的推薦系統：如規劃避免3D暈眩或撞到障礙物的虛擬和現實路徑、推薦畫面顯示內容以最大化社群共感和個人喜好、及虛擬世界社群網路中NFT交易推薦系統。 • 社群影響力分析與優化：如多面向社群影響力學習與預測、社群網路生成模型、具性別平等意識之影響力最大化、個人化密度彈性群體查詢與圖(graph)中子結構的資訊融合。 • 其他應用領域的推薦系統：如推薦系統多人毒害攻擊、為團體活動安排與活動潛在的參與者推薦。 (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。 • 虛擬實境和元宇宙網路：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，以最佳化多媒體網路傳輸效率及確保使用者的沉浸體驗。 • 行動邊緣運算網路：如結合數位雙生(digital twin)和分散式AI訓練架構，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，並採用真實資料集和AI模型驗證系統效能。 • 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。 元宇宙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。長期以來，我們關心未來元宇宙中的各種社群網路問題，包括虛擬實境的朋友和NFT推薦系統、即時串流平台推薦系統、社群影響力分析及社群資料探勘；此外，我們亦關心次世代網路優化問題，包括AI網路效能優化、有線及無線資源配置、單播/群播排程設計及社群物聯網(social IoT)和群眾外包(crowdsourcing)系統設計。在這裡，你將有機會學習到多項技術，包括圖神經網路、機器學習、生成模型、張量分解技術、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。歡迎有意出國留學、希望提升實作能力或對元宇宙創業充滿期待的同學，於今年夏天加入我們，一同探索未來元宇宙與AI網路的無盡可能。 A. Social network data mining, machine learning, and algorithm design: Research tensor decomposition, neural network, machine learning, and other technical solutions for: • Virtual, augmented reality (VR/AR) or metaverse recommendation system (e.g., user display configuration recommendation, planning a path avoiding 3D motion sickness and obstacles, and virtual world social network in NFT markets). • Social influence analysis and optimization (e.g., multi-channel influence diffusion model, generative models for social networks, fairness influence maximization, density personalized group query, and fusing graph substructures information into node features). • Recommendation systems for other applications (e.g., data poisoning attacks in multiplayer settings, group activities arrangement, and potential customer recommendations). 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. • Virtual reality (VR) and metaverse 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 optimize transmission efficiency and users' immersive experiences). • Mobile edge computing networks (e.g., incorporate digital twins and distributed AI architecture to build high-performance and reliable social IoT and crowdsourcing systems, and validate system performance via real AI models and datasets). • 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). Welcome those who plan to study abroad and enhance their implementation skills or are interested in the metaverse. Please join us this summer. Let's explore the opportunities of the metaverse and AI networking. |
PI個人首頁(PI's Information) : http://www.iis.sinica.edu.tw/pages/dnyang/ 實驗室網址(Research Information) : http://www.iis.sinica.edu.tw/pages/dnyang/ https:// Email : dnyang@iis.sinica.edu.tw; denianyang@gmail.com |
| 廖純中 Churn-Jung Liau | 應用邏輯 Applied Logic | 符號邏輯與應用，包括模態邏輯，知態邏輯，規範邏輯，多值邏輯，知識表徵與推理等。 We are interested in symbolic logic and its applications, including modal logic, epistemic logic, deontic logic, many-valued logic, knowledge representation and reasoning, etc. |
PI個人首頁(PI's Information) : http://www.iis.sinica.edu.tw/pages/liaucj/ 實驗室網址(Research Information) : https://chess.iis.sinica.edu.tw/lab/?cat=2 https:// Email : |
| 楊柏因 Bo-Yin Yang | 後量子密碼學 Postquantum Cryptography | 後量子密碼學是中大型量子電腦問世之後仍可保持安全性的公鑰密碼學。 本實驗室主要工作是在做後量子密碼學, 特別是其中的實作。 暑期研習的目標主要是動手做後量子密碼學, 如果暑假前即可開始最佳。 本實驗室不是 "量子密碼學", 想學量子密碼學的請出門右轉鐘楷閔教授研究室 Postquantum Cryptography (PQC) is the study of public-key cryptography that stays secure in the face of cryptographically relevant quantum computers. Our lab mostly does postquantum cryptography particularly PQC implementations. It is best if you can start before the summer and we aim to get hands-on experience implementing PQC. This is not the lab for Quantum Cryptography (QC), for QC please see Prof. Kai-Min Chung's lab. |
PI個人首頁(PI's Information) : http://www.iis.sinica.edu.tw/pages/byyang/ 實驗室網址(Research Information) : http://www.iis.sinica.edu.tw/pages/byyang/ http://www.iis.sinica.edu.tw/pages/byyang/ Email : byyang@iis.sinica.edu.tw |
| 修丕承 Pi-Cheng Hsiu | 可持續的微型機器學習 Sustainable TinyML | 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. |
PI個人首頁(PI's Information) : http://www.citi.sinica.edu.tw/pages/pchsiu/ 實驗室網址(Research Information) : https://emclab.citi.sinica.edu.tw/ https:// Email : pchsiu@citi.sinica.edu.tw |
| 林仁俊 Jen-Chun Lin | 可編輯的音樂驅動3D舞蹈創作 Editable Music-driven 3D Choreography | 編舞結合了技術與創意，要求動作序列與音樂完美匹配。在電影和遊戲產業中，創建3D舞蹈動畫不僅昂貴，而且耗時。這通常需要專業編舞設計動作、舞者精湛的執行，以及使用人體動作捕捉系統（MoCAP）的團隊來記錄這些動作，進而將它們應用於3D角色，以創造逼真的舞蹈動畫。這整個過程需要專業的舞蹈和編舞知識，以及對動作捕捉進行細緻的後期處理，如動作校正和編輯，從而顯著提高了製作原創3D舞蹈動畫的成本。因此，開發以數據為驅動的技術，即利用現有音樂與舞蹈動作資料來自動創造並允許用戶介入編輯的舞蹈動作生成系統，將顯著提高原創3D舞蹈動畫的製作效率並降低成本。在這項研究中，我們將發展以下技術：1. 微調大規模預訓練模型技術於跨媒體音樂到3D舞蹈映射，2. 擴散模型於音樂驅動的3D舞蹈創作，3. 文字引導的3D舞蹈生成與編輯技術。此外，我們還將探討基於照片/視訊的3D人體動作與形體估測，以及文字引導的人物3D動作與形體估測等研究議題。 實習生預計從上述議題或其他相關主題中選定題目進行研究。實習結束後，若同學表現優良則可繼續與實驗室合作研究並發表論文。 Choreography, an art that combines technique with creativity, requires the detailed design of movement sequences to align perfectly with music. In the film and gaming industries, creating 3D dance animations is both costly and time-consuming. This process typically involves choreographers designing the movements, skilled dancers performing them, and a team using a Motion Capture System (MoCAP) to record these movements for application to 3D characters, thereby creating realistic dance animations. Such a process necessitates expertise in dance and choreography and involves extensive post-processing of the captured movements, including motion correction and editing, which significantly raises the cost of producing original 3D dance animations. Hence, the development of data-driven technologies, which utilize existing music and dance movement data to automate and facilitate user-edited dance motion generation, will substantially improve efficiency and reduce costs in producing original 3D dance animations. Our research focuses on developing: 1. Advanced techniques for fine-tuning large-scale pre-trained models for cross-media music to 3D dance mapping, 2. Diffusion models for generating music-driven 3D dance, and 3. Techniques for text-guided 3D dance generation and editing. We will also delve into topics such as image/video-based 3D human pose and shape estimation, as well as text-guided 3D human pose and shape estimation. Interns are expected to select a research topic from the aforementioned topics or other related subjects. Upon completion of the internship, students with outstanding performance may continue to collaborate with the laboratory on research and publish papers. |
PI個人首頁(PI's Information) : http://www.iis.sinica.edu.tw/pages/jenchunlin/ 實驗室網址(Research Information) : https://sites.google.com/site/jenchunlin/ https:// Email : jenchunlin@iis.sinica.edu.tw |
| 吳真貞 Jan-Jan Wu | 深度學習於異質系統架構平台之資源配置與高效排程之技術研發 Resource Allocation and Operation Scheduling for Deep Neural Network Computing on Heterogeneous System Architectures | 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。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。本計畫所提之方法將可充分發揮各階層的平行度,包括資料平行度,管道平行度(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平行度(例如，充分發揮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. |
PI個人首頁(PI's Information) : http://www.iis.sinica.edu.tw/pages/wuj/ 實驗室網址(Research Information) : http://www.iis.sinica.edu.tw/pages/wuj/ https:// Email : wuj@iis.sinica.edu.tw |
| 張原豪 Yuan-Hao Chang | 基於記憶體與儲存體運算之設計、優化與模擬驗證 Design, Optimization, Simulation/Verification with In/Near Memory and Storage Computing | 這個專案旨在研究針對下一代量子模擬/驗證、機器學習演算法和神經網絡加速器的以內存為中心的計算架構。該專案的主題將集中在為目標應用選擇合適的以內存為中心的方法，並基於應用特性和底層內存/存儲計算硬件提出優化策略。我們將解決以下關鍵問題：（1）選擇和設計適當的以內存為中心的計算架構；（2）新興的非揮發性內存技術（例如 ReRAM、PCM、MRAM 和快閃記憶體），支持以內存（或存儲）進行計算；（3）基於目標應用和硬件設計特性的優化策略；以及（4）有效的模擬平台，用於估算所提出架構的效率。以下是該專案的一些關鍵問題和研究任務/目標： 以內存為中心的計算加速器：以內存為中心的計算的關鍵概念是通過減少數據移動的需求來消除數據傳輸的瓶頸。因此，選擇和設計目標應用的高效加速器的主要挑戰在於如何全面分析應用的特性和工作機制，數據移動以及需要仔細考慮的每個步驟的計算操作。 優化策略：儘管適當的硬件設計可以顯著提高性能，但僅通過硬件設計很難完全提高效率。因此，優化策略（例如量化、近似計算、剪枝、數據局部性和稀疏數據壓縮）應與所提出的硬件設計一起考慮。 任務1：利用處理-內存架構進行應用 處理-內存方法利用內存結構（例如忆阻器十字陣列）在內存讀操作期間處理數據並輸出結果。然而，當前的處理-內存方法僅支持簡單的計算操作，如乘-累加（MAC）和數據匹配。因此，更適合採用處理-內存方法的是那些需要大量並行簡單計算的應用，例如同時需要許多MAC操作的圖卷積網絡。此外，隨機森林算法必須並行處理多個需要簡單值比較操作的決策樹。將研究使用新興非揮發性內存技術（例如 ReRAM、PCM、MRAM）構建的十字陣列或內容寻址内存等處理-內存結構。 任務2：採用處理-鄰近-內存架構用於大數據應用 處理-鄰近-內存（PNM）將處理單元（如DPU）集成到內存模塊中，可以以更高的內部帶寬訪問數據，並執行比處理-內存更複雜的操作。對於需要複雜計算的應用，例如RNA-seq量化，處理-鄰近-內存方法更適合，因為很難設計一個能執行複雜操作的內存計算硬件。這類硬件的成本將過高，並限制內存的容量。這些方法的最關鍵問題是如何充分利用多個DPU和內部帶寬。至於潛在的應用，我們想研究硬件輔助的量子電路驗證，我們認為這可能非常適合處理-鄰近-內存方法。這種應用需要在經典計算機上運行的大量數據傳輸和乘法運算，所需的存儲空間可能對處理-內存方法來說不可行。然而，考慮到處理-鄰近-內存架構的算法設計將是一個挑戰。例如，可能需要重新設計在經典計算機上可行但在處理-鄰近-內存環境中並非最佳選擇的數據結構。 任務3：存儲內計算設計和優化 存儲內計算最適合在數據中心內運行的大數據應用。這些應用（例如量子電路模擬器、推薦系統或非結構化數據搜索）需要TB級別甚至PB級別的數據。我們可以朝兩個方向發展：（1）最先進的快閃磁碟通常配備一個處理單元來管理其輸入/輸出請求。處理單元通常相當強大，應該能夠處理更多的任務。我們希望利用這個計算資源作為存儲內計算（ISC）以優化我們應用的算法。此外，我們希望利用我們在驗證快閃記憶體固件方面的經驗，確保某些ISC設計的穩健性。（2）另一方面，從存儲中獲取大數據應用所需的所有數據是困難的。ISC可以在將數據發送到存儲之前使用內置的輕量級處理器（例如FPGA和SSD控制器）高效處理數據。我們還將探索在快閃記憶體讀取期間使用內存搜索功能的快閃記憶體，以在閱讀快閃記憶體時過濾不需要的數據的可能性。這有望成為一種有前途的解決方案。 This project aims to study the memory-centric computing architectures for next-generation quantum simulation/verification, machine learning algorithms, and neural network accelerators. This project's main theme will focus on selecting an appropriate memory-centric approach for a target application and proposing optimizing strategies based on the characteristics of the application and underlying memory/storage computing hardware. We will tackle the critical issues of selecting and designing (1) appropriate memory-centric computing architectures, (2) emerging non-volatile memory technologies (e.g., ReRAM, PCM, MRAM, and flash memory) that support in-memory (or in-storage) computation, (3) optimizing strategies based on the characteristics of the target application and hardware design, and (4) efficient and effective simulation platform to estimate the efficiency of the proposed architecture. In the following, we highlight some critical issues and research tasks/objectives of this project as follows: - Memory-centric computing accelerator: The crucial concept of memory-centric computing is to eliminate the bottleneck of data transferring by reducing the need for data movement. Hence the main challenge of selecting and designing an efficient accelerator for the target application falls on how to comprehensively analyze the application’s characteristics and working mechanisms, the data movement, and the computing operation of each step that needs to be carefully considered. - Optimizing strategy: Although a proper hardware design can dramatically improve performance, it is hard to fully enhance efficiency with hardware design only. Therefore, optimizing strategies (e.g., quantization, approximate computing, pruning, data locality, and sparse data compaction) should also be considered along with the proposed hardware design. Task 1: Exploiting the processing-in-memory architecture for applications The PIM approach utilizes memory structures such as the memristor crossbar array to process data during the memory read operation and output the result. However, the current PIM approaches only support simple computing operations like multiply-accumulate (MAC) and data matching. Therefore, it is more suitable to adopt the PIM approach for applications with a significant number of parallelly simple computations, e.g., graph convolutional networks that simultaneously need many MAC operations. Also, the random forest algorithm must parallelly process multiple decision trees with simple value comparison operations. PIM structures such as a crossbar or content-addressable memory built with emerging non-volatile memory technologies (e.g., ReRAM, PCM, MRAM) will be studied. Task 2: Adopting the processing-near-memory architecture for big-data applications PNM, which integrates processing units like DPU into the memory module, can access data with a higher inner bandwidth and perform more complicated operations than PIM. For applications that need complicated computation, such as RNA-seq Quantification, the PNM approach is more suitable because it is hard to design an in-memory computing hardware to perform the complicated operation. The hardware cost of such hardware will be too high and limit the capacity of the memory. The most critical issue of these approaches is how to fully utilize multiple DPUs and the inner bandwidth. As for potential applications, we would like to investigate hardware-assisted quantum circuit verification, which we believe might be a good fit for the PNM approach. Such an application requires enormous data movement and multiplication operations running on classical computers, and the required storage space might be infeasible for PIM. However, the algorithm design considering the PNM architecture would be a challenge. For example, one might need to redesign the data structures that are feasible in classical computers but are not the best choice considering the PNM environment. Task 3: In-storage computing design and optimization In-storage computing is most suitable for big-data applications that operate within the data center. These applications (e.g., quantum circuit simulators, recommender systems, or unstructured data searching) require TB-scale or even PB-scale data. We can take this direction in twofold. (1) The state-of-the-art flash disk usually comes with a processing unit to manage its input/output requests. The processing unit is often quite powerful and should be able to handle more tasks. We would like to exploit this computing resource as the in-storage computing (ISC) to optimize our application’s algorithms. Moreover, we want to use our past experience in the verification of flash’s firmware to ensure the robustness of some ISC designs. (2) On the other hand, it is hard to fetch all the data required by the big-data applications from storage to memory. ISC can efficiently process the data with an embedded light-weighted processor (e.g., FPGA and SSD controller) before sending data out of the storage. Instead of inserting a processing unit into the storage, we will also explore the possibility on utilizing flash memory with an in-memory searching function for filtering unwanted data during a flash memory read. This would potentially become a promising solution |
PI個人首頁(PI's Information) : https://www.iis.sinica.edu.tw/~johnson/ 實驗室網址(Research Information) : https://www.iis.sinica.edu.tw/~johnson//index_c.php https:// Email : johnson@iis.sinica.edu.tw |
| 吳廸融 Ti-Rong Wu | 深度強化式學習與電腦遊戲 Deep Reinforcement Learning and Computer Games | 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. |
PI個人首頁(PI's Information) : http://www.iis.sinica.edu.tw/pages/tirongwu/ 實驗室網址(Research Information) : https://github.com/rlglab https:// Email : tirongwu@iis.sinica.edu.tw |
| 王建堯 Chien Yao Wang | 多模態電腦視覺研究 Multimodal Research for Computer Vision | 電腦視覺技術在許多日常應用層面已臻成熟，人們開始將電腦視覺系統使用在一些高風險低容錯的應用情境，如自動駕駛。然而，純視覺解決方案在許多極端環境下無法提供可靠的視覺分析結果。此時，多種模態的訊號可以用來輔助電腦視覺任務完成低風險高容錯的預測結果。暑期實習的目標便是探討多模態資訊如何交互影響，使得深度學習模型，如圖神經網路等，能夠提取更值得信賴的預測結果。 Computer vision technology has matured in many daily applications, and people have begun to use computer vision systems in some high-risk, low-fault-tolerant application scenarios, such as autonomous driving. However, purely visual solutions fail to provide reliable visual analysis results in many extreme environments. At this time, signals of multiple modalities can be used to assist computer vision tasks in achieving low-risk and high-error-tolerant prediction results. The goal of the summer internship is to explore how multi-modal information interacts with each other so that deep learning models, such as graph neural networks, can extract more trustworthy prediction results. |
PI個人首頁(PI's Information) : http://www.iis.sinica.edu.tw/pages/kinyiu/ 實驗室網址(Research Information) : http://www.iis.sinica.edu.tw/pages/kinyiu/ http://www.iis.sinica.edu.tw/pages/kinyiu/ Email : kinyiu@iis.sinica.edu.tw |
| 廖弘源 Mark Liao | 多域電腦視覺研究 Multi-Domain Research for Computer Vision | 基於深度學習的電腦視覺技術是由資料驅動的方式進行模型訓練。因此，對於未知場域，深度學習往往會由於訓練資料的偏差而產生錯誤的判讀結果。暑期實習的目標是要探討如何利用深度學習方法，如圖神經網路，來加強深度學習模型的跨域推理能力。主要涵蓋多域的通用性與可調適性研究。 Deep learning-based computer vision technology conducts model training in a data-driven manner. Therefore, for unseen domains, deep learning often produces erroneous interpretation results due to bias in training data. The goal of the summer internship is to explore how to use deep learning methods, such as graph neural networks, to enhance the cross-domain reasoning capabilities of deep learning models. Mainly covers multi-domain generalization and adaptation research. |
PI個人首頁(PI's Information) : http://www.iis.sinica.edu.tw/pages/liao/ 實驗室網址(Research Information) : http://www.iis.sinica.edu.tw/pages/liao/ https:// Email : liao@iis.sinica.edu.tw |
| 洪鼎詠 Ding-Yong Hong | 深度學習軟體與硬體協同優化研究 Deep Learning Software/Hardware Co-optimization | 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, 優化深度學習程式, 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. |
PI個人首頁(PI's Information) : http://www.iis.sinica.edu.tw/pages/dyhong/ 實驗室網址(Research Information) : http://www.iis.sinica.edu.tw/pages/dyhong/ http://www.iis.sinica.edu.tw/pages/dyhong/ Email : dyhong@iis.sinica.edu.tw |
| 王釧茹 Chuan-Ju Wang | 基於檢索增強生成的大型語言模型（RAG-based LLMs）及其應用 Retrieval-augmented-generation-based (RAG-based) Large Language Models (LLMs) and its Applications | 研究主題將集中於找出關鍵或有趣的真實世界應用，並開發相應的基於檢索增強生成（RAG-based）的解決方案。這項研究可能涵蓋多個學科。例如，我們可能會探索非結構化的財務資料來建立金融問答系統，或者研究特定用途的應用程式，如為非通用程式語言提供的程式撰寫輔助系統。除了模型設計外，實習還將為參與者提供：1）親手體驗如何處理真實世界資料的機會；2）學習如何處理大規模資料並在Unix-like系統下進行系統化的實驗；3）通過前端網頁程式設計獲得視覺化成果的技能。 The research topics will focus on identifying critical or intriguing real-world applications and developing corresponding RAG-based (retrieval-augmented generation-based) solutions. This research may span multiple disciplines. For example, we might explore unstructured financial data to build a financial question-answer system or investigate specialized applications such as coding assistance systems for non-general-purpose programming languages. In addition to model design, the internship will offer participants the opportunity to 1) gain hands-on experience with real-world data, 2) learn how to handle large-scale data and conduct systematic experiments in Unix-like systems, and 3) acquire skills in visualizing outcomes using front-end web programming techniques. |
PI個人首頁(PI's Information) : http://www.citi.sinica.edu.tw/pages/cjwang/ 實驗室網址(Research Information) : http://cfda.csie.org https:// Email : cjwang@citi.sinica.edu.tw |
| 林仲彥 Chung-Yen Lin | 以人工智慧來解析生物醫學大數據 Harnessing Biomedical Big data in AI for Enhanced Quality of Life | 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），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分析，同時也定序、重組與註解了多個重要經濟生物之基因體。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，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。研究範圍以單細胞基因解析、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 is dedicated to unraveling the complexities of large-scale omics data, aiming to uncover the hidden secrets of biological regulation amidst the vast data landscape. We've integrated open-source tools with our in-house developed programs and platforms to successfully assemble, annotate, and decode numerous aquatic genomes of significant economic value. In parallel, we are incorporating AI methodologies into our metagenomic research and in the crafting of functional therapeutic peptides. A key objective is to innovate new methods to bridge the gaps in the human genome assembly, thereby laying a foundation for personalized and precision medicine. We plan to employ advanced techniques like deep learning to re-examine our studies and gain fresh insights. Our AI and biological expertise have culminated in the development of several platforms and applications, focusing on the smart typing of upper respiratory pathogens and the innovative identification and design of novel antibiotics. This multifaceted approach reflects our commitment to leveraging technology and biology for groundbreaking scientific advancements. |
PI個人首頁(PI's Information) : http://www.iis.sinica.edu.tw/pages/cylin/ 實驗室網址(Research Information) : http://eln.iis.sincia.edu.tw https://hub.docker.com/u/lsbnb Email : cylin@iis.sinica.edu.tw |
| 王志宇 Chih-Yu Wang | 邊緣智慧/量子網路 Edge Intelligence / Quantum Network | 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. |
PI個人首頁(PI's Information) : http://www.citi.sinica.edu.tw/pages/cywang/ 實驗室網址(Research Information) : http://snaclab.citi.sinica.edu.tw https:// Email : cywang@citi.sinica.edu.tw |
| 張佑榕 Ronald Y. Chang | Wireless-empowered AI Wireless-empowered AI | 見英文介紹 This internship will explore the intersection of wireless communication and machine learning, including (but not limited to): 1) Federated Learning via Over-the-Air Computation ("wireless for machine learning") 2) Reconfigurable Intelligent Surface (RIS) Assisted Communication 3) Federated Learning with Model Heterogeneity 4) Satellite-Enabled Federated Learning Ideal candidates should have good programming skills and have interest or experience in one or more of the following: 1) Wireless Communication: multiple-input multiple-output (MIMO) communication, cell-free networks, reconfigurable intelligent surface (RIS) assisted communication, satellite communication, etc. 2) Machine Learning: convolutional neural networks (CNNs), graph neural networks (GNNs), reinforcement learning (RL), federated learning (FL), 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 possible upon demonstration of satisfactory performance and mutual agreement. |
PI個人首頁(PI's Information) : http://www.citi.sinica.edu.tw/pages/rchang/ 實驗室網址(Research Information) : https://www.citi.sinica.edu.tw/~rchang/ https:// Email : rchang@citi.sinica.edu.tw |
| 王新民 Hsin-Min Wang | 語音辨識、語者暨語言辨識、語音合成與轉換、語者分離與分段、語音機器翻譯與問答系統 speech recognition, speaker/language recognition, speech synthesis and conversion, speaker separation and segmentation, speech translation, spoken question answering system | 我們實驗室致力於符合我國語言使用情境（國語、臺語、客語、原住民語、英語）的語音處理研究，學術研究與系統開發兼重。 1）在語音辨識方面，我們的辨識器要聽得懂年輕人的國語和老人家的臺語、客語、原住民語，對於喜歡繞英文的人來說，也要難不倒它；加上語者過濾的技術，更要能專一辨識出特定語者的聲音，達到客製化、不受環境干擾的效果。 2）在語音合成方面，我們的合成系統不僅要會說國語、臺語、客語、原住民語，也要能夠使用語音轉換的技術，客製出使用者指定的人聲，不僅有娛樂性，更是語音保存、有聲書製作不可或缺的技術。 3）目前的「純」機器翻譯及「純」文字問答並不稀奇，真正能便利地應用於智慧家庭、音箱的使用環境還得藉助語音。因此，不僅我們的翻譯系統要能處理國人所需的國臺、國客、國英互譯，我們的問答系統也要達到快速聆聽、高效反應的效果。 4）針對各種的受損語音，例如全喉切除患者的語音，我們希望利用語音處理技術去提升語音品質及可理解度。 Our laboratory is dedicated to research on speech processing in line with the context of language use in our country (Mandarin, Taiwanese, Hakka, aboriginal languages, and English), with emphasis on both academic research and system development. 1) In terms of speech recognition, our recognizer must be able to recognize Mandarin of young people and Taiwanese, Hakka, and aboriginal languages of the elderly. For those who like to mix English in their speech, it should not be troublesome. Coupled with the technology of adaptation, it is necessary to be able to identify the voice of a specific speaker specifically to achieve the effect of customization and robustness to environmental interference. 2) In terms of speech synthesis, our synthesis system must not only speak Mandarin, Taiwanese, Hakka, and aboriginal languages, but also be able to use voice conversion technology to customize the user-specified human voice, which is not only entertaining, but also an indispensable technology for spoken language preservation and audiobook production. 3) The current text machine translation and text Q&A are not uncommon. It is really convenient to use in smart homes and speaker environments with the help of voice. Therefore, not only our machine translation system must be able to handle the translation between Mandarin and Taiwanese, Mandarin and Hakka, Mandarin and aboriginal languages, and Mandarin and English that local people need, but also our Q&A system must achieve the effect of fast listening and efficient response. 4) For various types of damaged speech, such as the speech of total laryngectomy patients, we hope to use speech processing technology to improve speech quality and intelligibility. |
PI個人首頁(PI's Information) : https://www.iis.sinica.edu.tw/pages/whm/ 實驗室網址(Research Information) : https://slam.iis.sinica.edu.tw/ https:// Email : whm@iis.sinica.edu.tw |
| 呂俊賢 Chun-Shien Lu | AI與深度學習的安全與隱私 Security and Privacy in AI and Deep Learning | 研究關於AI與深度學習的安全與隱私的各項議題, 包括 1. Adversarial Attack and Defense 2. Deepfake Detection 3. Backdoor Attack and Defense 4. Privacy Protection and De-identification 5. Deep Learning Hashing and Adversarial Hashing 6. 或以上之外的其它相關議題! 研究關於AI與深度學習的安全與隱私的各項議題, 包括 1. Adversarial Attack and Defense 2. Deepfake Detection 3. Backdoor Attack and Defense 4. Privacy Protection and De-identification 5. Deep Learning Hashing and Adversarial Hashing 6. 或以上之外的其它相關議題! |
PI個人首頁(PI's Information) : http://www.iis.sinica.edu.tw/pages/lcs/ 實驗室網址(Research Information) : http://www.iis.sinica.edu.tw/pages/lcs/ https:// Email : lcs@iis.sinica.edu.tw |
| 蔡懷寬 Huai-Kuang Tsai | 生物資訊 Bioinformatics | 近年來，生物資訊在醫學領域受到相當大的關注，不僅是定序技術的進步，或是在數據分析的能力也持續在更新。本實驗室除了與國內許多生物試驗單位合作外，更從世界知名大型研究機構取得資料，以大數據分析解開生物中複雜的調控機制與原理。 我們的研究方向著重於真核生物的基因體，並整合多種體學資料 (multi-omics) ，以不同的生物觀點，更具系统性的探討基因體層級的交互關係，與其在演化上的重要性。而我們最新的研究主題著重在人類重大疾病與大數據資料庫整併，藉由不同的序列資料尋找目前生物醫學上未解出的困境。我們的研究方法透過資料探勘和機器學習來建立分析模型，用來預測生物的基因調控，同時達到個人化精準醫療的開發及癌症預測應用。 本實驗室想要找對生物資料運用感興趣的大專生,你可以來自資工或生物背景，但應該熟悉至少一種程式語言及對生物學感興趣。我們會提供生物資訊學相關領域的知識訓練，因此只要您對於跨領域研究感興趣，也想解決目前生物領域面臨的瓶頸，歡迎您加入我們的研究團隊。 The Tsai lab studies big data from biological systems using bioinformatic techniques and statistical methods. We work with biologists to seek insights into the genomics of eukaryotic organisms. By integrating multi-omics data, we study genome-wide regulatory systems on gene expressions and their significance in evolution. In addition, we are currently expanding into the area of biomedical informatics, aiming at integrating disease information with sequencing data for development of applications in precision medicine. We use methods such as data mining and machine learning in our studies on regulatory mechanisms in genomics, with the aim of building predictive models with potentials for applications. We are seeking interns with a background in either computer science or biological science. The applicant should have experience in using at least one programming language and have a strong interest in biology. We will provide training in bioinformatics-related domain knowledge, and we expect our interns to be able to learn from team members from different backgrounds. If you are passionate in taking up the challenge of solving biological problems with techniques in informatics, we welcome you to join our team! |
PI個人首頁(PI's Information) : http://www.iis.sinica.edu.tw/pages/hktsai/ 實驗室網址(Research Information) : https://bits.iis.sinica.edu.tw/?id=1 https:// Email : hktsai@iis.sinica.edu.tw |
| 呂及人 Chi-Jen Lu | 深度學習的原理與應用 Deep learning: foundations and applications | 研究深度學習的原理，並探索深度學習在強化學習、影像處理、自然語言等各個領域的應用。 Study the foundation of deep learning, and explore its diverse applications in various areas such as reinforcement learning, computer vision and natural language processing. |
PI個人首頁(PI's Information) : http://www.iis.sinica.edu.tw/pages/cjlu/ 實驗室網址(Research Information) : http://www.iis.sinica.edu.tw/pages/cjlu/ https:// Email : cjlu@iis.sinica.edu.tw |
| 陳駿丞 Jun-Cheng Chen | 基於多模態條件之影像生成與偽造偵測 Generation and Detection of Image Synthesis based on Multi-modal Conditions | Based on diffusion model-generated digital content, the quality of the generated content has been significantly improved. Models such as ControlNet and Multi-ControlNet further allow users to control the content generated by the diffusion model through different modal conditions. This research aims to guide interns in familiarizing and mastering the background knowledge of relevant generative models, and further improve model performance, such as addressing issues of artifacts (flaws in generated images), bias (whether the generation of boys or girls tends towards a specific ethnic group), and accuracy (whether the generated content meets the given conditions). On the other hand, the study explores how to design detectors to effectively distinguish whether images are AI synthesized. Outstanding performers in the summer internship will have the opportunity to be employed as part-time research assistants after the internship ends, with chances to publish in well-known conferences or journals. 基於擴散模型生成的數位內容，其生成內容品質已經大幅提升, 而ControlNet及Multi-ControlNet等模型，進一步允許使用者透過不同模態的條件對擴散模型生成的內容進行控制，本研究目標帶領實習生熟悉與掌握相關生成模型背景知識，並進一步改善模型性能，如破圖 (生成圖片的瑕疵) 、偏見 (生成的男孩或女孩是否傾向特定族群)、正確性(生成的內容是否符合給定得條件)。另一方面，針對生成的內容，研究如何設計偵測器以有效辨別圖片是否為AI合成。 暑期實習表現優良者，實習結束後可獲聘兼任研究助理繼續研究，並且有發表知名會議或期刊論文的機會。 |
PI個人首頁(PI's Information) : http://www.citi.sinica.edu.tw/pages/pullpull/ 實驗室網址(Research Information) : http://www.citi.sinica.edu.tw/pages/pullpull/ https:// Email : pullpull@citi.sinica.edu.tw |
| 穆信成 Shin-Cheng Mu | 函數程式語言的推導與證明 Reasoning about Functional Programs | 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 concurrent 程式的型別系統與命令式語言(imperative languages)的推理。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。 本年度我們可探索的幾個題目包括 * 迴圈與累積參數。如何將遞迴函數轉成單一迴圈？以 quicksort 為例，如果要用一個迴圈做 quicksort, 我們需要另用一個資料結構暫存待計算的資料。但一般說來，這個資料結構該如何設計？我們已知這可能和累積參數 (accumulating parameter) 與 continuation 有關，但仍有許多細節待研究。 * 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？ * 資料結構與 Catalan 數：Catalan 數 (C n) 是組合數學中時常出現的數字。它同時是含 n+1 個節點的「三叉樹」的個數、含有 n 組括號而語法正確的字串的數目、n+2 邊型分割成三角形的方式的個數... 是否能將這些種種不同情況表達成依值資料結構，用 Catalan 數做 index, 將它們統合起來？ * 或然率單子 (monad) 與 binomial 特性。將一個正面機率為 p 的不公平硬幣丟 n 次，出現 i 個正面的機率恰好是其二項式係數。這個性質能不能用或然率單子的性質證明出來？這將是一個研究或然率單子的練習。 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。 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。 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: * Constructing loops from recursive programs. To perform quicksort using a single loop, for example, we have to use an auxiliary data structure to store the sublists to be sorted. But how to design such data structures in general? It is speculated that accumulating parameters or continuations play a role, but a lot remain to be studied. * 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? * Catalan number and data structure. A Catalan number is many things: the number of trivalent trees with n+1 nodes, the number of ways to construct well-formed parentheses, the number of ways to divide a polygon with n+2 sides into triangles... Can the relationship between these cases be made more clear if we index data structures using Catalan numbers? * Probabilistic monad. Given an unfair coin with probability p, the possibility of having i heads after n tosses is exactly its binomial coefficient. Can we prove it using properties of probabilistic monad? This will be an exercise to help us getting more familiar with probabilistic monads. 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. |
PI個人首頁(PI's Information) : http://www.iis.sinica.edu.tw/pages/scm/ 實驗室網址(Research Information) : https://scm.iis.sinica.edu.tw/ncs/ https:// Email : scm@iis.sinica.edu.tw |
| 劉庭祿 Tyng-Luh Liu | 基於生成式建模的電腦視覺技術 Computer Vision Techniques via Generative Modeling | We are interested in developing generative techniques that leverage diffusion models (or additionally interacting with language/vison foundation models) to address a broad range of computer vison applications, including image/video representation learning, efficient sampling, image/video editing, anomaly/deepfake detection, visual reasoning, and robotics. We are interested in developing generative techniques that leverage diffusion models (or additionally interacting with language/vison foundation models) to address a broad range of computer vison applications, including image/video representation learning, efficient sampling, image/video editing, anomaly/deepfake detection, visual reasoning, and robotics. |
PI個人首頁(PI's Information) : http://www.iis.sinica.edu.tw/pages/liutyng/ 實驗室網址(Research Information) : https://homepage.iis.sinica.edu.tw/~liutyng/ https:// Email : liutyng@iis.sinica.edu.tw |
| 柯向上 Hsiang-Shang ‘Josh’ Ko | 附正確性證明的函式程式與型式化的數學基礎 Verified functional programming and formalised foundations of mathematics | 每當我們寫出一個數學定理和證明，其實也就寫出一個具有型別（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，隨後衍生出眾多有成熟實作的證明輔助器（proof assistants）和依值型別程式語言（dependently typed programming languages），我們現在寫的證明和程式已融為一體，而且電腦能夠提供更多資訊協助我們寫出對的證明／程式。 實習的焦點會放在 Agda 這個程式語言學界常用的語言，由基本的 Agda 程式寫作開始，依興趣可選擇往兩個方向延伸：一個方向是 (Homotopy) Type Theory 這套新興的數學基礎，探索這套新數學語言如何重塑數學理論；另一個方向是證明與程式合一的依值型別程式寫作 (dependently typed programming)，由電腦提供豐富型別資訊協助寫出正確程式，不需另寫太多證明。 實習型式類似讀書會，自行研讀材料和動手實作，並定期與老師同學們分享討論；若有餘裕也可做個小專案。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），除非已獲得偶數年修業證書，或有其他特殊情況可與指導老師商量。 請在自傳內敘述為何對此主題有興趣。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。 參考讀物：請見英文介紹最後。 Whenever we write down a mathematical theorem and its proof, we have also written down a 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 Type Theory is a common foundation for mathematics and programming, and has spawned numerous proof assistants and dependently typed programming languages, in which proofs and programs take the same unified form, and the computer can provide more information to help us develop correct proofs/programs. We will focus on Agda, which is a popular language in the programming languages research community. Starting with basic programming in Agda, we can explore either (Homotopy) Type Theory, a newly developed foundation of mathematics that can be used to reformulate mathematical theories, or dependently typed programming, where proofs are embedded into programs, allowing the computer to provide detailed type information to help us write correct programs while avoiding separate proofs. The 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. 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; either signing up for credits or just sitting in). Exceptions apply if you have already obtained a certificate for passing the course in an even-numbered year or if there are special circumstances, which should be discussed with the supervisor. Please explain why you are interested in this research topic in your statement of purpose. If you are applying for a FLOLAC special internship (see the FLOLAC website https://flolac.iis.sinica.edu.tw for more detail), please make the intention explicit in the statement of purpose, and provide concrete reasons why you are suitable for working in this research topic, so that your application can be prioritised. # References ## (Homotopy) Type Theory Martín Hötzel Escardó [2019]. Introduction to univalent foundations of mathematics with Agda. DOI: 10.48550/arXiv.1911.00580. https://www.cs.bham.ac.uk/~mhe/HoTT-UF-in-Agda-Lecture-Notes/ Simon Thompson [1999]. Type Theory and Functional Programming. Addison-Wesley. ISBN: 9798482847145. https://www.cs.kent.ac.uk/people/staff/sjt/TTFP/ ## Dependently typed programming in Agda Ana Bove and Peter Dybjer [2009]. Dependent types at work. In International LerNet ALFA Summer School on Language Engineering and Rigorous Software Development 2008, volume 5520 of Lecture Notes in Computer Science, pages 57–99. Springer. DOI: 10.1007/978-3-642-03153-3_2. https://www.cse.chalmers.se/~peterd/papers/DependentTypesAtWork.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 [2023]. The power and joy of abstraction. https://josh-hs-ko.github.io/blog/0034/ Conor McBride [2011]. Ornamental algebras, algebraic ornaments. https://personal.cis.strath.ac.uk/conor.mcbride/pub/OAAO/LitOrn.pdf |
PI個人首頁(PI's Information) : https://www.iis.sinica.edu.tw/pages/joshko/ 實驗室網址(Research Information) : https://josh-hs-ko.github.io https:// Email : joshko@iis.sinica.edu.tw |
| 王建民 Chien-Min Wang | 機器學習與遺傳式編程 Machine Learning and Genetic Programming | (1) 人機組隊之深度強化學習：人機組隊 (Human-Autonomy Teaming, HAT) 已成為最新興的 AI 究趨勢之一，包括以人為中心的人智計算，和深度強化學習 (Deep Reinforcement Learning, DRL) 的自治 AI 算法。先進的 DRL 系統除了可達到智慧系統與人類進行更密切的合作外，同時亦可作為人類最佳的模範幫手、教練或競爭伙伴，以執行更合乎道德規範的互動性組隊，進而完成更合理且更適用的目標任務。HAT 基於人機之間的共享權限，以正確地學習共通指令、共同目標和競爭伙伴關係的模型；本研究成果將輔助 HAT 成為更有效的決策系統，同時達到高相容性和可靠性的人機系統。本研究計劃旨在構建一個具有交互運作、協作團隊和風險分析集合的仿真人 DRL 系統。在研究計畫中提出的創新方法，可作為未來 HAT 與 DRL 系統開發的重要基礎，以實現未來動態和自主環境中，更重要、與人相容、直觀和可靠的自主系統。 (2) 使用遺傳式編程探究監督式機器學習：本研究計畫透過嘗試解決兩個不同需求的應用問題，來探討監督式機器學習的兩個不同階段。第一個應用問題是要找出最能符合、解釋觀察樣本資料的機率分佈數學模型，此與監督式機器學習的第一階段(訓練/學習)目標一致；而第二個應用問題(網際服務品質時間序列預測)則要求模型除了要能符合學習/訓練資料之外，還需具有一般性的能力，以便未來能正確地應對未曾見過之資料或情況，此與監督式機器學習第二階段對模型的要求相同。此外，不同於時下熱門的深度學習方法使用類神經網路模型和倒傳遞式訓練，本研究計畫探索機器學習的另一種可能性與方向，也就是遺傳式編程 (Genetic Programming, GP) 。其使用數學表達式模型和演化式搜尋學習，有益於機器學習結果的理解、推導與運用，符合 Explainable AI 所提倡之概念。 (1) Deep Reinforcement Learning for Human-Autonomy Teaming： Human-Autonomy Teaming (HAT) has become one of the most emerging AI research trends consisting of Human-Centered Computing and self-governed AI algorithms such as Deep Reinforcement Learning (DRL). The advanced DRL system with sophisticated design allows intelligent gent's closer cooperation with humans while performing moral, reasonable, and applicable tasks as humans' most exemplary assistants, tutors, and/or competition partners. Based on HAT's pursuing the collective goals of sharing the authority, offering instructions, and/or competitions between humans and machines, the research outcomes will help HATs become more effective decision-support systems while sustaining a highly compatible and reliable Human-AI system. Furthermore, this research proposal aims at constructing a human-level DRL system with an interactive, collaborative teaming and risk analysis integration. The novel approach proposed in this study can enhance and extend the development and important foundation of future HAT with DRL systems as Explainable AI methodology for more considerable, human-compatible, intuitive, and reliable applications in future dynamic and autonomous environments. (2) Exploring Supervised Machine learning with Genetic Programming： Through solving two individual application problems, this research proposal investigates two separate stages of supervised machine learning. First, the main purpose of the former application problem is to identify the probability distribution function for a set of observation data, which matches the goal of the training/learning phase of supervised machine learning. Afterwards, the second application problem concentrates on Web service QoS time series prediction, which requires the generated model to be capable of dealing with unseen data or situations rather than just merely fitting provided data. Moreover, instead of adopting the widely used deep learning techniques, this research proposal tries another possibility and research direction, i.e., genetic programming, which employs evolutionary searching/learning strategy and mathematical expression-based models that are helpful for the understanding and use of the outcomes of machine learning process. |
PI個人首頁(PI's Information) : http://www.iis.sinica.edu.tw/pages/cmwang/ 實驗室網址(Research Information) : http://www.iis.sinica.edu.tw/page/research/ComputerSystem.html?lang=zh https:// Email : iho@iis.sinica.edu.tw |
| 蘇黎 Li Su | 以人為核心的音樂人工智慧 Human-centered music AI | 音樂與文化科技實驗室致力於探討最前沿的數位訊號處理與深度學習技術，應用於各種音樂人工智慧的熱門議題。我們特別關注以人為核心的音樂人工智慧研究，包括但不限於： 1. 音樂內容辨識與理解、自動採譜 2. 音樂與多媒體內容之生成 3. 計算音樂學 4. 深度學習、認知科學與生物音樂學 5. 音樂人工智慧技術之應用與評估方法 我們歡迎資訊/電機等理工相關科系或音樂相關科系背景，或有志於跨領域研究之同學應徵。熟悉深度學習、音訊處理、影像處理、電腦圖學、認知科學、音樂學等任一領域者優先考慮。 The mission of the Music and Culture Technology Lab is to solve cutting-edge research topics on music AI using novel deep learning and signal processing research technologies. We specifically concentrate on of human-centered music AI, including (but not limited to) the following research directions: 1. Music content understanding and automatic music transcription 2. Music and multimedia content generation 3. Computational musicology 4. Deep learning, cognitive science and biomusicology 5. Application and evaluation method of music AI technology We welcome students with background in EE/CS, musicology and related fields and who are interested in inter-disciplinary research to join our intern program. The students who are familiar to deep learning, signal processing, image processing, computer graphics, cognitive science, or musicology will be considered with first priority. |
PI個人首頁(PI's Information) : http://www.iis.sinica.edu.tw/pages/lisu/ 實驗室網址(Research Information) : https://sites.google.com/view/mctl/home https://github.com/Music-and-Culture-Technology-Lab Email : lisu@iis.sinica.edu.tw |
| 陳伶志 Ling-Jyh Chen | 從智慧城市到幸福城市：使用微感測器與智慧優網探究城市的未來發展 Making Smart City a Happy City: using low-cost sensors and AIoT | 智慧城市現已成為近年來都市發展的重要趨勢。但隨著愈來愈多的資訊科技融入我們的城市生活，大家也開始思考：這些新興的建設是否真的達到了預期的目標？這些智慧科技究竟能否實質提升我們城市的生活品質？我們認為，城市是一個將物理環境與社會經濟融合在一起的複雜系統。因此，城市發展不應只是盲目追求智慧化，更應該著重於提升都市生活的方式與節奏，從而增進居民的整體幸福感。 在這個計畫中，我們希望能透過具有高時間和空間解析度的微感測物聯網，結合學術理論、創新思維及實際應用價值，進行資料的融合和深入分析。本計畫以人為核心，將重點放在民眾的生活步調、對環境的感知，以及空間認知等方面。透過運用微型環境感測器，結合人工智慧與時空大數據分析，我們將致力於探索幸福城市的衡量指標和推動策略。我們熱烈歡迎對物聯網、人工智慧、微感測、以及大數據分析懷抱熱情的夥伴加入我們的行列，共同參與這段旅程。讓我們一起帶來創新的智慧城市感測想法與挑戰，為提升城市居民的幸福感而努力。 Smart cities have emerged as a key trend in urban development in recent years. However, as urban life increasingly incorporates information technologies, there is growing contemplation about whether these new developments truly fulfill their intended goals. People are questioning whether these innovative technologies enhance the quality of city life. We advocate that a city is a multifaceted system, intertwining the physical environment with the social economy. Urban development should not just chase technological sophistication but instead aim to enrich the quality and rhythm of urban living, thereby promoting the well-being of its inhabitants. Our project combines data mashup and sophisticated analysis with academic theory, creativity, and practical value. This will be achieved through the micro-sensing Internet of Things, characterized by high-resolution temporal and spatial data. Our human-centered approach focuses on the pace of life, environmental awareness, and spatial perception of individuals. By employing micro-environmental sensors and integrating artificial intelligence with spatio-temporal big data analysis, we aspire to identify and enhance the metrics and strategies for creating a happier city. We invite partners interested in the Internet of Things, artificial intelligence, micro-sensing, and big data analysis to join us in this venture. We will learn, collaborate, enjoy the process, and strive for meaningful outcomes. For any inquiries, please don't hesitate to contact us. |
PI個人首頁(PI's Information) : https://cclljj.github.io/ 實驗室網址(Research Information) : https://cclljj.github.io/research/ https:// Email : cclljj@iis.sinica.edu.tw |
| 王柏堯 Bow-Yaw Wang | 密碼程式形式驗證工具 Formal verification tools for cryptographic programs | 暑期實習期間，我們將利用形式方法驗證密碼程式之正確性，提出改進工具之方法並進行實作。 We will use a tool to verify cryptographic programs formally, discuss how to improve the tool, and implement new features on the tool. |
PI個人首頁(PI's Information) : http://www.iis.sinica.edu.tw/pages/bywang/ 實驗室網址(Research Information) : http://www.iis.sinica.edu.tw/~bywang https:// Email : bywang@iis.sinica.edu.tw |
| 王有德 Yu Te Wang | 腦機介面、人機互動、混合實境 Brain-Computer Interface, Human-Computer Interaction, Mixed Reality | 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、用大腦打字。 目標： (1) 實作高頻的SSVEP（穩態視覺誘發電位）視覺刺激於BCI系統中。 (2) 改編和設計最先進的解碼算法，以提高多模生物信號解碼的準確性。 (3) 整合並開發大腦輸入應用程式。 二、為XR設備建立BCI系統原型。 本專案將專注於開發可攜式BCI系統，目標如下： (1)利用3D列印，製造測量生物信號的感應器。 (2) 設計和製作一款頭戴式裝置，與現成的XR設備整合（例如HTC Vive、Microsoft HoloLens2或Meta Quest2）。 (3) 開發一個BCI應用程式。 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。 Welcome to the Brain-Computer Interface (BCI) world! This summer, we are looking for students who are interested in BCI-enabled mixed reality (XR) devices. 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. Project description: 1) Typing by brain. This project has three aims: (1) Implement a high frequency SSVEP (steady-state visual evoked potential) visual stimulus for a BCI system. (2) Adapt and design state-of-art decoding algorithms to improve decoding accuracy of multi-modal bio-signal. (3) Integrate and develop an end-to-end brain-typing application. 2) Prototyping a BCI system for XR devices. This project will focus on the development of a portable BCI system. There are three aims in this internship (1) 3D-printing the sensors for measuring the bio-signal. (2) Design and prototype a headset to integrate with off-the-shelf XR devices (ex, HTC Vive, Microsoft HoloLens2, or Meta Quest2). (3) Develop an end-to-end BCI application. 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. |
PI個人首頁(PI's Information) : http://www.citi.sinica.edu.tw/pages/yutewang/ 實驗室網址(Research Information) : http://www.citi.sinica.edu.tw/pages/yutewang/ https:// Email : yutewang@citi.sinica.edu.tw |
| 李育杰 Yuh-Jye Lee | 基於零信任架構的聯邦式學習 Zero-Trust Federated Learning | 聯邦式學習(Federated Learning; FL)是一種分散式的機器學習架構，使用戶端能夠在不揭露隱私資料、不共享數據的情況下共同進行去中心化的機器學習。然而，FL仍有許多安全漏洞。我們的目標是基於零信任架構(Zero-Trust Architecture)，發展可信賴且安全的學習方式，來提升FL的安全性、隱私性、及可信度。 在這項研究中，我們將探討： (1) 在可能有資料異質性的情況下，透過分散式最佳化來訓練FL模型 (2) 運用差分隱私(Differential Privacy), 零知識證明(Zero-Knowledge Proof), 安全多方計算(Multi-Party Computation), 同態加密(Homomorphic Encryption), 異常偵測(Anomaly Detection)等資安手段來建構零信任的聯邦式學習 Federated Learning (FL) is a cutting-edge distributed Machine Learning (ML) framework that enables various clients to conduct decentralized ML without exposing the clients’ private data or violating data sharing issues. However, the FL training process is susceptible to several security vulnerabilities. We aim to develop a trustworthy and robust learning system by incorporating Zero-Trust Architecture, thereby enhancing the security, privacy, and trustworthiness of the FL framework. We will start with the linear/nonlinear support vector machines, random forests, or regression models and extend to deep neural networks. In this research, we investigate: (1) training the model by solving a distributed optimization problem, which poses its own set of challenges regarding heterogeneous data (2) applying Differential Privacy (DP), Zero-Knowledge Proof (ZKP), Multi-Party Computation (MPC), Homomorphic Encryption, Anomaly Detection, and other computer security techniques in to build the Zero-Trust FL Architecture |
PI個人首頁(PI's Information) : http://www.citi.sinica.edu.tw/pages/yuh-jye/ 實驗室網址(Research Information) : http://www.citi.sinica.edu.tw/pages/yuh-jye/ https:// Email : yuh-jye@citi.sinica.edu.tw |
| 王大為 Da-Wei Wang | 醫療資料分析應用與民眾自主意願表達之資訊架構設計 Medical data analysis and the information framework of personal data autonomy | 我們的研究興趣包含2個主題，其一是醫療資料分析應用將考慮使用醫療資料進行跨域資料分析。另一個是希望透過資訊技術的協助，設計並實現健康資料治理機制，讓民眾對於自己的個人資料使用的自主意願表達更容易執行，進而提升民眾對於資料使用的參與度，也可以讓當事人知道資料的使用情形，獲得更多的資訊回饋，促進資料利用的良性循環。 Our research interests include 2 topics. One is that using statistical data analysis methods and machine learning technology for cross-domain data research. The other is to use the stance of information technology to design and implement a health data governance mechanism to make it easier for people to express their autonomy of personal data. Thereby, individual participation in the use of data can be increased and promotes a virtuous cycle of data utilization. |
PI個人首頁(PI's Information) : http://www.iis.sinica.edu.tw/pages/wdw/ 實驗室網址(Research Information) : https://chess.iis.sinica.edu.tw/lab/ https:// Email : |
| 葉彌妍 Mi-Yen Yeh | 聯邦學習與生成式AI應用 Federated Learning and Generative AI Applications | 本實驗室目前幾個研究方向如下: 1. 生成式AI應用:大型語言模型興盛，我們研究相關應用包括(1)如何於提示工程中結合知識圖譜於非特定大型語言模，讓語言模型能有更精確的回答、(2)如何使機器學會讓語言模型忘記特定的資訊以保護特定隱私、(3)利用語言模型協助自動生成目標導向對話等。 2. 聯邦式學習: 巨量資料在不同的端點設備大量且快速產生，因此促成了去中心化學習的架構。我們深究在這樣的學習架構下依據不同的運算條件限制，針對不同模型與不同學習任務做各式的評估指標優化。 We have the following research directions: 1. Generative AI applications: Large language models (LLM) are booming, and our research on related applications includes (1) how to combine knowledge graphs with non-specific large language models in prompt engineering to enable language models to answer more accurately, (2) how to ask machines to “unlearn” in LLM applications, that is, to make language models forget specific information to protect specific privacy, and (3) use language models to help automatically generate goal-oriented conversations. 2. Federated learning: Big amounts of data are generated in large quantities and quickly across different endpoint devices, thus enabling the architecture of decentralized learning. Under such a learning architecture, we conduct various evaluation index optimizations for different models and learning tasks according to various computing conditions. |
PI個人首頁(PI's Information) : http://www.iis.sinica.edu.tw/pages/miyen/ 實驗室網址(Research Information) : http://www.iis.sinica.edu.tw/pages/miyen/ https:// Email : miyen@iis.sinica.edu.tw |
| 陳郁方 Yu-Fang Chen | 自動化形式化驗證相關研究 Topics on formal verification related research | 我們的研究室專注於形式化驗證，這是一種強大的工具，用於確保軟體和硬體系統的正確性。我們的研究涵蓋多個方向，包括但不限於： - SMT（可滿足性模理論） 探討SMT在解決複雜系統中的應用，如自動驗證、模型檢查等。 -系統程式驗證 研究作業系統、編譯器和其他系統軟體的形式化驗證方法。 -量子程式驗證 推進在量子計算領域的形式化驗證技術，解決相應的挑戰。 實習內容: 作為我們的實習生，您將參與以下活動之一或多個： -閱讀並分析相關的學術論文，掌握形式化驗證的最新進展。 -參與小型研究項目，挑戰實際的形式化驗證問題。 -參與研討會和學術交流，分享您的發現並與其他研究者互動。 Our lab is dedicated to formal verification, a powerful tool for ensuring the correctness of software and hardware systems. Our research spans various directions, including but not limited to: -SMT (Satisfiability Modulo Theory) Explore the applications of SMT in solving complex system verification problems, such as automated verification and model checking. -System Program Verification Investigate formal verification methods for operating systems, compilers, and other system software. - Quantum Program Verification Advance formal verification techniques in the field of quantum computing, addressing corresponding challenges. 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. |
PI個人首頁(PI's Information) : http://www.iis.sinica.edu.tw/~yfc 實驗室網址(Research Information) : http://www.iis.sinica.edu.tw/~yfc https:// Email : yfc@iis.sinica.edu.tw |
| 鄭湘筠 Hsiang-Yun Cheng | 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 | 近年來大數據分析 (深度學習、圖論分析、基因序列分析等) 逐漸盛行，這些應用程式在運算時往往仰賴高效的巨量資料存取，然而目前主流的運算系統無法滿足這些需求，使得我們必須重新思考如何設計未來的電腦系統。 其中一個極具潛力的設計方向是從傳統以運算單元為主的系統切換到以記憶體為主的運算系統，藉由在記憶體內或週邊直接做部分的運算避免資料傳輸造成的效能瓶頸。許多新興的記憶體技術，像是ReRAM、PCM、MRAM、FeFETT、NAND/NOR Flash等，兼具存儲與運算功能，能在記憶體陣列內實現許多運算(例如矩陣向量乘法、位元邏輯運算、向量相似度搜索等)，為實現以記憶體為中心的運算系統帶來新的曙光，產業界也積極開發於3D堆疊式記憶體(3D-stacked memory)或DRAM晶片上加入簡單運算單元之技術以實現近記憶體運算 (例如Samsung HBM-PIM及AxDIMM、SK Hynix AiM、UPMEM PIM等)，但由於硬體技術尚不成熟以及和傳統截然不同的運算模式，且不同大數據分析之演算法也具有不同之運算與資料存取特性，在系統設計上有許多尚待克服之挑戰。 本實習計畫的目標為針對大數據分析之各式應用情境，探討不同層面上之設計挑戰，包括電路與元件階層、計算結構階層、及演算法階層，並以軟硬體協同設計的方式，充分發掘以記憶體為中心之優勢，設計高效能低耗電之新世代運算系統。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，以記憶體內資料特徵分析的方式，實現高效能低耗電之深度學習。 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，如圖學習、推薦系統、基因序列比對等，設計異質性記憶體為中心運算系統。 3. 透過演算法與硬體協同設計，以跨記憶體層級之記憶體內運算，提升大規模數據分析，如圖學習、資料樣式探勘等，之運算能源效率。 In recent years, data analytics applications, including deep learning, graph analytics, and genome data analytics, which necessitate the processing of substantial data volumes, have gained significant popularity. These big data applications require efficient data access. Unfortunately, mainstream computing systems are not tailored to meet these requirements, compelling us to reconsider the fundamental design of future computing platforms. One promising solution is to transition from a contemporary processor-centric design to a revolutionary memory-centric approach. Unlike current systems that suffer from energy-inefficient data transfers between separate compute and memory/storage units, memory-centric systems perform computations directly in or near memory/storage units. Emerging memory technologies like ReRAM, PCM, MRAM、FeFET、NAND/NOR Flash, etc., enable various computations (e.g., matrix-vector multiplication, bit-wise logic operations, vector similarity search) to be parallelly executed directly in memory arrays. Leading industry vendors are also developing techniques to integrate simple computing units in 3D-stacked memory or DRAM DIMMs, enabling near-memory computing (e.g., Samsung HBM-PIM & AxDIMM, SK Hynix AiM, UPMEM PIM, etc.). Despite its promise, bringing such a system into practice remains challenging due to hardware constraints and the distinct computing characteristics of various applications. Our goal is to investigate design challenges across various system layers, encompassing device/circuit, architecture, and algorithm levels. We aim to propose cross-layer designs that fully capitalize on the potential of in-memory/near-memory computing systems. Candidate topics include, but are not limited to, the following: 1. Energy-efficient deep learning through in-memory/in-storage feature analysis 2. Design heterogeneous memory-centric computing systems for applications with irregular data accesses and composite program behavior (e.g., recommendation system, genomic sequence analysis, graph learning, etc.) 3. Algorithm-hardware co-design to enable cross-layer memory-centric computing for large-scale data analytics (e.g., large-scale graph learning, pattern mining, etc.) |
PI個人首頁(PI's Information) : http://www.citi.sinica.edu.tw/pages/hycheng/ 實驗室網址(Research Information) : http://www.citi.sinica.edu.tw/pages/hycheng/ https:// Email : |
| 曹昱 Yu Tsao | 基於AI的生理醫學聲學訊號處理 AI-based Biomedical Acoustic Signal Processing | 我們的研究主題是基於人工智慧的生醫聲學訊號處理，我們主要處理的訊號包括心電圖（ECG）、肌電圖（sEMG）、脈搏波形（PPG）、心肺音 (Lung Sound and Heart Sound) 以及語音訊號 (Speech Signals)。我們的目標是開發創新的演算法，以處理這些訊號中的雜訊，並有效地分析或擷取這些生理訊號中所包含的重要資訊，以實現遠程健康狀況診斷、病情評估和康復進展追蹤。 這個研究方向結合了人工智慧和生醫領域的專業知識，是一個嶄新且對人類健康非常有益的研究主題。我們期待著能夠在這個領域取得令人矚目的成果，為遠程監測和健康管理提供更加革命性的解決方案。 Our research topic is artificial intelligence (AI) -driven biomedical acoustic signal processing. The primary signals we deal with include electrocardiograms (ECG), surface electromyograms (sEMG), photoplethysmogram waveforms (PPG), lung sounds, heart sounds, and speech signals. Our objective is to develop innovative algorithms to address noise within these signals and effectively analyze or extract crucial information embedded in these physiological signals, enabling remote health condition diagnosis, disease assessment, and rehabilitation progress tracking. This research direction integrates expertise from both AI and biomedical fields, making it a novel and highly beneficial research topic for human health. We look forward to achieving significant advancements in this field, providing groundbreaking solutions for remote monitoring and healthcare management. |
PI個人首頁(PI's Information) : http://www.citi.sinica.edu.tw/pages/yu.tsao/ 實驗室網址(Research Information) : https://bio-asplab.citi.sinica.edu.tw/ https:// Email : yu.tsao@citi.sinica.edu.tw |
| 蔡孟宗 Meng-Tsung Tsai | 串流式圖論演算法 Graph Streaming Algorithms | 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。 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 a graph problem can be computed using O(n) space, where n denotes the number of vertices in the input graph. We assume that the edges of the input graph are given to algorithms one by one, in an arbitrary order, and only once. Note that an n-vertex 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. 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? |
PI個人首頁(PI's Information) : http://www.iis.sinica.edu.tw/pages/mttsai/ 實驗室網址(Research Information) : http://www.iis.sinica.edu.tw/pages/mttsai/ https:// Email : mttsai@iis.sinica.edu.tw |
| 陳亮廷 Liang-Ting Chen | 程式語言與數學的基礎 Foundations for Programming Languages and Mathematics | # 背景說明 程式設計與建構式數學本質上兩者互相呼應。更精確地說，以型別論（type theory）為基礎設計的程式語言可以用來論證數學事實。反過來，建構式數學的基礎（foundation of mathematics）可以當作程式語言。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⋯⋯等諸如此類的聯繫。我們將此邏輯與計算之間的聯繫稱作為 Curry-Howard 對稱性。 # 實習主題 在暑期實習期間，我們會一起探索程式語言的計算與邏輯特性，並試著在以下主題之間完成一個短期的研究專題： 1. 用建構式基礎的形式化某種數學理論 2. 函數式程式的驗證 3. 程式語言的數學模型 在實習開始兩週的時間，實習生需要先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)。可正式修課取得學分或旁聽參與課程，但須參與考試。接著我們會一起選讀論文，並從以上主題間選出一個研究問題。在接下來的時間會在我的指導下探討該研究問題。 # 必要條件 申請者需要一定的數學成熟度，以及對計算理論或邏輯有基本認識。參與此暑期實習，請在申請信中至少包含這兩點的解釋。 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。 # 關於我 一般來說我對於電腦科學中的數學結構特別感興趣，尤其是在理論電腦科學中被稱為「理論 B」的分類。我的研究工作主要用邏輯與範疇論方法應用在依值型別程式設計、程式語言、邏輯與自動機理論等領域。 近來透過 Curry-Howard 對稱，我主力於應用型別論的計算以及邏輯性。並以定理證明器 Agda 作為工具，在程式語言的理論與實務上都有所產出。以此方法我們可以得到得到驗證，可以重現的研究成果。 # Background Programming and constructive mathematics are inherently interconnected. More precisely, a programming language based on type theory, without any use of general recursion, can be used to reason about mathematical facts. On the other hand, a foundation of constructive mathematics can serve as a programming language where logical propositions can be viewed as types, proofs as programs, and proof-checking as type-checking. This intersection of logic and computation is called as the Curry-Howard correspondence. # Potential topics During the summer internship, we will together explore the computational and logical facets of programming languages. The focus will be on undertaking short-term research projects within the following themes: 1. Formalising some theory in mathematics or computer science, 2. Verification of functional programs, or 3. Mathematical models for programming languages. Interns are expected to attend the Formosan Summer School on Logic, Language, and Computation (FLOLAC) at the National Taiwan University during the first two weeks, either for credit or as auditors. An exam will follow at the end of FLOALC. (Unless a candidate is able to demonstrate sufficient understanding of the topics covered at FLOLAC.) Subsequently, interns will have two to three weeks to read relevant papers, identify a research problem within the specified themes, and then undertake an in-depth investigation under supervision. # Requirements Prospective candidates should have a certain level of mathematical maturity and a basic understanding of programming or logic. Please explain why you are qualified for the internship in your statement of purpose. If you are applying for a FLOLAC special internship (URL), please make the intention explicit in the statement of purpose, and provide concrete reasons why you are suitable for working in this research topic, so that your application can be prioritised. Applicants are required to explain their qualifications for the internship in their statement of purpose. Those applying for a FLOLAC special internship should explicitly express their intention in the statement and provide concrete reasons showcasing their suitability for the research topic. # About me I am generally interested in mathematical structures in computer science, specifically, the so-called ‘theory B’ of theoretical computer science. My work so far spans dependently typed programming, programming languages, logic, and automata theory, employing logical and categorical methods. Recently, I have been exploring the computational and logical aspects of type theory, particularly through the lens of the lens of Curry-Howard correspondence. I have worked on both the theory and practice of programming languages using Agda, a proof assistant and a dependently typed language. This approach allows me to produce reliable, reproducible, and sometimes useful results. |
PI個人首頁(PI's Information) : https://l-tchen.github.io 實驗室網址(Research Information) : https://l-tchen.github.io https:// Email : liang.ting.chen.tw@gmail.com |
| 莊庭瑞 Tyng-Ruey Chuang | 1. 相互豐富的研究資料與維基資料 2. 創新以及可永續的研究資料管理和協作 1. Mutual Enrichment between Research Data and Wikidata 2. Innovative and Sustainable Research Data Management and Collaboration | 1. 開放研究資料已不再是新的口號。從 Open Data 到 FAIR Data 有各種倡議與原則，但面臨研究實務上的議題時，針對不同學科領域的研究資料，存在著不同程度的想像空間及挑戰。好的研究，始於好的資料。除了使用 Wikidata 做為「研究資料寄存所」 (網址： https://data.depositar.io/about ) 的資料集關鍵字來源，以加強資料集之間的語意連結之外，本實驗室也陸續嘗試與不同學科領域的研究夥伴，進行各種資料的爬梳及結構化的處理。 2. 我們將從社群 (community)、技術 (technology)、協作 (collaboration)、以及研究 (research) 四面向， 協力發展台灣本地在研究資料管理的實踐社群。此實踐社群將以我們已開發的「研究資料寄存所」（網址： https://data.depositar.io/about ）為實踐的場域之一。本計畫的預期成果包括：培養研究資料管理人才、參與開放資料軟體系統的國際協作專案、提昇研究資料管理實踐社群在台灣的規模與內涵、以及參與並貢獻所能到全球研究社群。 更多資訊可參閱以下： 1. "Openness" https://policyreview.info/glossary/openness 2. Datasets and presentations from the depositar project team: https://data.depositar.io/organization/depositar 1. We will study WiIkidata, and use Wikidata to enrich research datatsets, and vice versa. We will further enhance our research data repository (called depositar, website: https://data.depositar.io/about ) with Wikidata, and vice versa. 2. We will work on the community, technology, collaboration, and research aspects of research data management. We will help develop a community of practice for research data management in Taiwan. A research data repository we have developed (called depositar, website: https://data.depositar.io/about ) can function as a starting place where the communities practice research data management. The expected outcome of such a effort includes: cultivating research data management talents, participating in international collaborative projects for open data software systems, elevating the scale and capacity of the research data management community in Taiwan, and participating in and contributing to the global research community. Please refer to the following for more information: 1. "Openness" https://policyreview.info/glossary/openness 2. Datasets and presentations from the depositar project team: https://data.depositar.io/organization/depositar |
PI個人首頁(PI's Information) : http://www.iis.sinica.edu.tw/~trc/ 實驗室網址(Research Information) : https://lab.depositar.io https://rdm.depositar.io Email : trc@iis.sinica.edu.tw |
| 黃彥男 Yennun Huang | (1)入侵偵測與防禦韌性研究 (2)人工智慧與資料中心能源管理 (3)資料安全隱私保護 (4)沉浸式數位建模設計舒適居住空間之研究 (1)Intrusion Detection and Defense Robustness Research (2)Artificial Intelligence & Data Center Energy Management (3)Data Security Privacy Technology (4)Research the Persuasive Power of Digital Immersive Modeling for Space-Sufficient Housing | (1)入侵偵測與防禦韌性研究 - 入侵偵測系統研發與測試 - 對抗例生成技術研發 - 自動化入侵規則生成 - 人工智慧框架使用 - 系統稽核紀錄和網路封包分析 (2)人工智慧與資料中心能源管理 本計畫提出資料中心運行能源效率優化方案，通過整合這些技術，發展節能減碳技術，達成淨零減排的目的，使資料中心更節能，更具韌性和適應性，滿足現代運算需求的高要求，研究項目如下。 -開發一種 AI 模型 (深度學習與機器學習) ，根據輸入的工作負載預測溫度和能源消耗，透過量測硬體的老舊程度、工作負載、溫度和能源消耗等資訊，分析彼此之間的關係。 -實施基於 AI 的 HVAC(Heating, Ventilation, and Air Conditioning)系統控制，系統控制和溫度調節機器人 (無人載具) 。 -分析並優化虛擬化和雲端運算的使用以實現更好的能源管理，設計一種線上排程機制以處理機器學習/高效能運算(High Performance Computing, HPC)工作負載。 (3)資料安全隱私保護 - 從事資料安全與隱私保護之相關領域研究工作 - Python、Linux、Matlab、Deep Learning Toolbox基礎程式撰寫能力 - Deep Learning、Federated Learning、Homomorphic Encryption技術相關研究 (4)沉浸式數位建模設計舒適居住空間之研究 本研究計畫重點分析如何利用沉浸式科技減少居住空間的浪費，已提高人們的居住品質。通過提供充足的虛擬居住和休閒體驗，本研究將探究如何使小坪數住宅仍能被人們認為是合適且裡想的居住空間。針對本項計畫所面臨到的問題以下三點做研究探討： - 建立一個擁有用戶社群的元宇宙平台，已創建虛擬居住環境。 - 識別元宇宙中針對用戶數據和隱私的各類攻擊者，並建立更安全、更可靠的NFT交易系統。 - 提供多樣化的沉浸式虛擬展覽、互動展演等XR內容，將知識轉譯納入元宇宙之中，使得觀者可以以元宇宙的形式接觸吸收新知。 (1)Intrusion Detection and Defense Robustness Research - intrusion detection system development and testing - adversarial sample generation - automatic intrusion rule generation - AI framework usage - system audit logs and network packets analysis (2)Artificial Intelligence & Data Center Energy Management This project proposes an energy efficiency optimization solution for data centers. Using these technologies, we aim to achieve net-zero emissions, make data centers more energy-efficient, resilient, and adaptable, and meet modern computing demands. Included in the research are: -Developing an AI model (using deep learning and machine learning techniques) to predict temperature and energy consumption based on input workloads. We analyze the relationships between them by measuring factors such as hardware, workload, temperature, and energy consumption. -Implementing AI-based HVAC (Heating, Ventilation, and Air Conditioning) system control, including system control and temperature regulation robots (autonomous vehicles) . -Analyzing and optimizing virtualization and cloud computing for better energy management: Designing an online scheduling mechanism to handle machine learning/high-performance computing (HPC) workloads. (3)Data Security Privacy Technology - Engaged in research work in related fields of data security and privacy - Basic programming ability for Python, Linux, Matlab, Deep Learning Toolbox - Research on deep learning, Federated learning, and Homomorphic Encryption technology (4)Research the Persuasive Power of Digital Immersive Modeling for Space-Sufficient Housing This project focuses on analyzing how immersive technology can be used to reduce the waste of housing space and improve the living space quality of human beings. By providing adequate virtual housing and leisure experiences, this study aims to find out how a smaller-scaled living space can still be considered appropriate and desirable by residents. There are 3 sub-projects included: - Creating a virtual housing environment by building up a metaverse platform with a user community. - Identifying different kinds of attackers on user data and privacy in the metaverse. Building a safer and more reliable NFT trading system. - XR content curation: provide diverse immersive/interactive virtual exhibitions for users to absorb knowledge and enjoy their leisure time. |
PI個人首頁(PI's Information) : http://www.citi.sinica.edu.tw/pages/yennunhuang/ 實驗室網址(Research Information) : http://www.citi.sinica.edu.tw/pages/yennunhuang/ https:// Email : yenjoanna@gmail.com |
| 馬偉雲 Wei-Yun Ma | 大型語言模型的記憶力研究 - 封閉型生成式QA Memory Research of Large Language Models - Closed Generative QA | 新的知識不斷的發生，使得大型語言模型(LLMs)在實際應用上多半是搭配檢索來使用，即先檢索相關知識再一併交由LLMs來生成。也就是目前流行的RAG (Retrieval-Augmented Generation)技術。不過，最近有研究者，包含Google，開啟了另一條擷取知識的方向 - 封閉型生成式QA，他們把語料庫中的所有資訊都編碼進LLMs的參數裡，讓LLMs型能夠掌握盡可能多的知識，讓LLMs不需要Retrieval，而能夠直接給出答案。我們可以做個類比：RAG就像是open-book的QA，而封閉型生成式QA就像是closed-book的QA。 封閉型生成式QA的困難點是LLMs的記憶力不足，甚至常常記錯或產生幻覺。研究人員對LLMs的記憶力機制也未完全掌握。這兩年來，對於大型語言模型的記憶力方面的研究開始多了起來，如Meta和Google都有許多針對大型語言模型記憶力的調查，例如：資料集大小、學習率和模型大小對記憶的影響，發現更大的語言模型在所有設定下記憶訓練資料的速度更快等等發現。 我們過去有開發大型語言模型的豐富經驗。事實上，第一個繁體中文優化的LLM - Bloom-zh系列即出自本實驗室與聯發科以及國教院的合作。目前我們正著手研究Llama2的記憶力，目的是打造新的架構或是訓練方法來增進記憶力，並應用在封閉型生成式QA。 在今年暑假，我們也開放數個名額開放給實習生，一起參與這個有趣又有挑戰的研究。 New knowledge continually emerges, leading to the predominant use of Large Language Models (LLMs) in practical applications with retrieval, wherein relevant knowledge is first retrieved and then processed by LLMs for generation. This approach is known as the popular Retrieval-Augmented Generation (RAG) technique. However, recent researchers, including Google, have embarked on another avenue of knowledge extraction - Closed Generative QA. They encode all information from the corpus into the parameters of LLMs, enabling LLMs to acquire as much knowledge as possible without the need for retrieval, allowing them to provide answers directly. We can draw an analogy: RAG is like open-book QA, while Closed-Generative QA is akin to closed-book QA. The challenge in Closed Generative QA lies in the limited memory capacity of LLMs, often resulting in memory errors or hallucinations. Researchers have not fully understood the memory mechanism of LLMs. Over the past two years, research on the memory aspect of large language models has seen increased attention, with Meta and Google conducting numerous investigations related to LLM memory, such as dataset size, learning rates, and the impact of model size on memory. They discovered findings like larger language models memorize data faster under all settings. We have a rich history of developing large language models. In fact, the first Traditional Chinese-optimized LLM - the Bloom-zh series - originated from our collaboration with MediaTek and the National Academy for Educational Research. Currently, we are researching the memory of Llama2 with the aim of creating new architectures or training methods to enhance memory and apply them to Closed Generative QA. This summer, we are opening several internship positions for individuals interested in participating in this exciting and challenging research. |
PI個人首頁(PI's Information) : https://homepage.iis.sinica.edu.tw/pages/ma/index_zh.html 實驗室網址(Research Information) : https://ckip.iis.sinica.edu.tw/ https:// Email : ma@iis.sinica.edu.tw |