---
title: Seminar
---
# 1102 Seminar
[TOC]
## 02/25 AI & chip

> **EDA for AI** & **AI for EDA**
### NNeed: An **E**valuation and Design Space **E**xploration Engine for C**NN** Accelerators using Extensive **D**ataflow Analysis
> [name=國立陽明交通大學電子工程學系 黃俊達教授]
- Architecture Developmemt & Algorithm Refinememt for Intelligent Computing
- EDA for AI
#### Introduction
- Biological NN -> Artificial Neural Network (ANN), MLPerception (MLP), Convolution NN, Recurrent NN
- Accelerator
- SoC : high performance, low power
- two PEs:
- **Fine-grained** architecture
- high flex
- simple
- high regularity and capacity
- **Coarse-grained** architecture
- Few complex PEs
- High utilization
- Low power and small area
- Systolic-Array based
#### Motivation and Previous Work
- Accelerator Design
- **HUGE** design space
- Need an ultra-eff egine to speed up the whole eval process
- eval egine
- practical hw
- effi dataflow
- rich ana
- fast turnaround time
(CNN model, config., constraints->ana egine->perform power area, hw arch.)
- simulation vs. analysis
- trade-off (speed / accuracy)
#### Methology
- which one is constraint? weight or input?
- **given constraint, then it could give you a best solution**
#### Hardware Template
- PE array
- Global buffer
- DRAM
- Input Stationary
- Weight Stationary
- Output Stationary
- Tiles and Tasks
- task is a tile-level computation job
- Task Sequence Example
- External Data Manhrmrny
- Traditional Dataflows
- Dataflow and Data Reuse
- IS, WS, OW is only be consider from PEs, so in general, there are $3!$ possible computation (I/W/O)
- there are two kinds of tasks: data-bound, compute-bound
- compute-bound: a better PE
- data-bound: higher DRAM bandwidth

#### Experiment Result
- depend on what you want
- In general, owi/oiw perform better then others, but it is **in generel**, different results in different layers (data will be smaller after long-term training, but weights are going to be extremely large)
- performance curve is **NOT** symmetric, but at first we will give it a square shape (**NOT** always the best)

### When AI meet EDA (Electronic Design Automation) -- 以類比設計自動化為例
> [name=國立陽明交通大學電機系 劉建男教授兼系主任]
- 國內電子設計翹楚!! (面不改色說著幹話🍵)
- character: 一本正經說幹話, 應該是面癱
- VSLI/CAD (電腦輔助設計)
- AI for EDA
#### Why EDA is important in IC design
- IC 設計最上游產業 (積體電路設計最重要的軟體)
- 精密儀器跟EDA軟體的需求(中美貿易戰如何打壓hauwei)
- CAD Research in VLSI Design
- more productivity -> CAD (Computer-Aided Design)
- exist everywhere
#### What CS can do in EDA
- synthesis = automation + optimization
- optimize for performance
- optimize for yield
#### Machine Learning for everything
- AI in EDA Tools: data analysis and management + ML + optimization
#### **ML-assisted evolu algorithm for circuit sizing**
- ML is used for predicting the design (after evolutionary operation)
- DNN is better than SVM (too simple)
#### conclusion
- ML + para = speed up at least 20 times
- EA + DNN + Force Model(their design) -> yield 97.5%

---

## 03/04 ZYXEL 合勤科技產品事業、人資經理
### 產業
> 產品事業管理部 葉宗昀(Gavin Yeh) 資深經理
- What are they doing ?
- ADSL、電信業、5G、資安
- 防火牆、網路設計代工
- 軍事武器訊號技術(ex. 雷達)
- R&D 人才培育
- 疫情下的市場趨勢
- 網路已變必需品
- IoT Connected devices 依賴度提高
- 利用 Internet 提升商品附加價值
- Opportunity
- E-Commerce
- Digital Payment
- Digital Media Spend
- Social Media
- Threatens
- Cyberattacks
- Ransomware Extortion
- Fake News/ Information
- Fraud
> 危機就是轉機
- Security
- Telco 電信
- Fiber 成本下降、收益提升(網速越快收費越高)
- Average Billing per user Increasing
- Operating Expense Decreasing
- Fixed-Mobile Convergence Strategy
- 綑綁消費
:::success
找國內外上市公司的財報，了解公司分析的未來趨勢及發展方向
（順便看股票好不好😄，以便未來投資
:::
- 軟體開發需求
- 知識財: 手機的作業系統是價值所在
- 比起會某一種程式語言，注重基本邏輯能力、學習能力
- 主要開發 **Middle Ware** base on Linux kernel
- Solution 調整
- 瑪斯克的Backhaul
### 人資
> 全球人資暨行政處 黃惠玲(Feeling Huang) 資深經理
-
賣服務

---
## 03/11 Poly 數位學習教育方案(產品發表)
> 賴巍政經理
- Polycom & Plantronic
- NASA 指定耳機
- 服務需求，問題解決
- 遠距教學
- 軟體跨平台
- 鏡頭追蹤: 講者、移動、多人。
- 有多種教學模式，因應故種教學方式
- 2秒是人體舒適切換的速度
- 收音: 高清、定位人聲、防ECHO、降低環境噪音
- Speaker tracking 聲波(束波)定位
- 利用多個麥克風所收到聲音的時間差來計算角度
- 使用至少三個麥克風矩陣，做數波定位
- 噪音屏蔽AI、虛擬隔音牆。來屏蔽外界的聲音，只收人聲
- 根據頻率來判斷人聲，屏蔽其他噪音

---
## 03/18 On the Overlapping Community Detection ProblemFile
> 彭勝龍
> 台北商業大學創意科技與產業設計系主任
### Introduction
- Social Networks
- Node: Individual, Organization
- Edge: Relation, Interaction
- Communities
- A group of vertices having a high density of edges, and a lower density of edges between different groups.
- Clusters in PPI Networks
- PPI: Protein-Protein Interaction
- a cluster can form a gene team or be involved in a pathway
> cluster: 群聚效果
- Community Detection
- aims to discover a community structure
- Community Structures(**C**)
- Disjoint
- Overlapping(close to the truth)
- Quality Functions
- Q fuction: dishoint
- 
- EQ funcition: overlapping
- 
> 兩 **C** 必較 Q value: 越高越好
- 用來分析並比較分群演算法優劣
### A Neighbor Melding Algorithm
- Neighbor Fusion for two **C**
- node 具有高影響力者可加入該群 **C**
> deg(u) >= max{E,E+1}
> degree: node 相連數量
- 重複融合直至 Quality 沒有改善
- 

### Hierarchical Merging Algorithms
- It shows an evolution of communities in a social network.
- It can be treated as a trend of community integration.
- Binary Tree
- Spilt(up to down) or Merge(down to up)
- A General Bottom-Up Approach
- 

#### EAGLE Algorithm
- Maximal Cliques
- Find out all basic communities
- Compute similarity
#### Our Algorithm
- Blocks of close neighborhoods
- a maximal biconnected component of a graph
### Coucluting Result
- Are th quality function good enough?
- A partition without any cut get Q = 0.083
- How to measure an evolution tree for predicting a trend of community integration?

----
## 03/25 產業AI化 - 台灣人工智慧學校
> [台灣人工智慧學校 官網](https://aiacademy.tw/)
> 蔡明順 校務長
- Outside-in 從需求的角度看資工專業的潛力
- 智慧路燈(含wifi等功能)
- 
- 人工智慧產業化
- 科技產業是現代最重要的公司
- 產業轉型(石油 --> 金融 --> 軟硬體科技)
- 軟體吃掉硬體，AI吃掉軟體
- AI是主流，其他像是元宇宙之類的，只是AI的附屬
- AI已被各國當成國家戰略
- 先進製造業、量子資訊科學、5G/B5G、AI是US政府投資的四逛關鍵技術
- 
- 數據、場景、演算法、半導體晶片與？？是AI重要的關鍵
- 現代進入Software2.0
- 
- Internet拉平了「資訊量」， AI拉平了「智能量」的差距
- AI產業化（AI+） vs. 產業AI化（+AI）
### 產業應用
- 智慧製造瑕疵檢測
- Edge AI
- AOI (Automated optical inspection)
- 智慧醫療 Prognostic and Health Management
- 避免誤診
- 仍在初期
- 智慧交通 - 智慧駕駛
- 
- 智慧金融
- AI發展程度，現在大概發展到level2、3
- 資訊安全未來需求很大
- [MLOps](https://en.wikipedia.org/wiki/MLOps)
### Question
- 您剛剛特別以德系和美系來區分外商公司，請問他們有甚麼不一樣的地方?
- Edge AI是拿來加速的嗎?
- Data Enginner vs Data Scientist
- 產業 AI 諮詢? 產業有需求 我們握有技術但不了解 請問現在有什麼平台協助二方面的協調嗎

---
## 04/01 YOLOv4 及其產業應用
> [name=中央研究院 廖弘源 特聘研究員/所長]
> 人工智慧學校副校長
> IEEE fellow

### 人工智慧在音樂上的應用
- Using AI automatically sequence the song
- to complete music vedio
e.g. 演唱會會有 serverl fans record and official version，用 AI turn them into a complete vedio

#### Issues needed to be addressed
- 各個影片觀看順序 - visual storytelling
- The diffrent of smartphone visual/audio quality
- language of flim

#### Potential Issue
- How to present the emotion, ideas, and art of a music director into a concert video mashup process?
- 符合音樂導演要求和經驗

#### Solution
- Visual
- Shot Classification based on Error-Weighted Deep Cross-Correlation Model
- Define the six camera shots by distance
- basically a pattern matching problem + correlation
- EW-Deep-CCM 17% error remain, sepcific in the simple shot (cause too many shot changes)
- Input: 1000+ category of objects, video stream from adele concert
- Process: Fuse the two using a Deep Neural network
- Ouput: 8 categories of stage shot
- Conditional Random ield-based (CRF)Apporach1.
- + CRF -> 90% :+1:

### Seethevoice: Learning from Music to visual storytelling of shots
- 用音樂驅動並直接選取適當的視訊片段
(Use AI pick a shot by music audio)

### Problem
- Must achieve real time response, which shot type to adopt

### System Framework
- KD
- Train a teacher net and a student net, teacher net
- Knowledge Distillation
- He who teaches, learns
- The student info turn back to teach teacher net, then teacher distill again till the model stable.

### Yolov4
- The world’s most **accurate and speedy** object detector during (Sept. 28 – Dec. 15, 2020)
- Smart City Traffic Flow Solutions
- Edging computing but use limited device ( TX2 )
- final goal: reinforcement learning -> 控制交通號誌
- object detection is a must (and Yolov4 is a detection model)
- traffic detection 貢獻之一: 轉向bounding box (YOLO V4 default is 正向)

### Questions
- Q1. What is the difference between the purpose for two prediction (Dense prediction, Sparse prediction)
- Q2. Occlusion problem
- Multiple Object Tracking

## 04/15 Clustered-based Multi-pin Fine Pitch Ball Grid Array Substrate Routing Optimization
> 劉一宇 台科大資工系 副教授
- 晶片設計
- 硬體描述語言
- EDA
- FBGA (Fine-pitch Ball Grid Array)
- Substrate: 晶片載板
- Routing: 繞線，避免線之間接觸導致短路。
- via: 在不同層電路間的樓梯
### Introduction
- Fine-pitch: 距離約等於 8~10 根頭髮(外面那些銀白色球體的間距)
- Structure
- Flip-chip
- 繞線技術越精細、成本越高
- Via 打孔受限於成本，size 為線體的 4~6 倍
- finger 的部分採用 any angle
- 手機製造
- 晶片: 10W
- 電壓: 1V
- 電流: 10A
### ILP(Interger Linear Programming)
- 任何要連接的兩點(起點finger --> 終點球體)
- keep-out zone(for finger)

    in --> (node) --> out
    1                  1

- Constraints
- in + out <= 2
- out - in
- 1: 起點
- -1: 終點
- 0: 過路的或是雜魚
- Relaxation
### Multi-pin Net Issue
- 前面討論的都是 two-pin，兩個點之間連接
- Algorithm
1. Design
2. Grouping
3. Spanning Tree(Delaunay Triangulation)
4. Connection
- Intra: finger to finger, bump ball to bump ball
- Inter: finger to bump ball
> 遠距離就交給 ILP 處理
5. Layout

---
## 04/15 Metaverse 元宇宙
> 佐臻 company
> AR Glasses
> New Brain-Computing Interface

## 04/29 我在「幸福企業」- 10年來學到的事情
> 王鐘逸 研究員

p.32 學長在進入業界後，有沒有遺憾當初在學期間沒有去做到的事情?或是如今學長看著底下帶的新人，有覺得我們需要多去努力的部分?

- 英文能力
- 快速抓到重點
ps. 不過他也覺得溝通很重要

Q: 那請問學長會用甚麼方式來練習溝通? 除了平常報告之外。

## 05/20 量子霸權與物聯網時代之密碼策略
> 許宏誌 助理教授
> 量子(電腦)簡介：與傳統bit不同，一個位元可以同時存在0 or 1，測量時，0 & 1會依據他的機率出現。 (我這樣講不知道你聽不聽的懂)
> 博士論文: High-capacity and Efficient-Embedding Steganography

### Threats
#### Quantum Supremacy
- **Shor's Algorithm**: 質因數分解演算法
- Threats to public-key crypto-systems
- **Input**: an odd composite number with n bits
- **Output**: a non-trival factorization of **Input** with some probability.
- Quantum parallelism(量子位元之間的糾纏)
- Reduce to order-finding problem
- 用在 **RSA**
- Time complexity: O(n^3logn)

- **Grover's Algorithm**: 無序資料庫搜索演算法
- Security-level shrinkage of private-key crypto-systems
- Search database with probabilities.
- O(N) -> O(sqrt(N))
- The security level of AES128 will shrink from 2^256 to 2^128.
- 

#### Internet of Things (IOT)
- Side-Channel Attack for private-key crypto-systems
- **SCA 旁道攻擊**
- 
- 幫忙縮短時間

### Solutions
- For quantum computing
- public-key crpto-systems
- Post quantum cryptography
- private-key crpto-systems
- Extending the key length of AES
- For side-channel attack in IoT
- Customized AES
- Signal masking
- Pipeline

> Candidate: 需支援七種加解密方式

#### NP-hard
- ex. 離散對數問題
- **Non-deterministic Polynomial**
- [什麼是P, NP, NP-Complete, NP-Hard問題](https://www.ycc.idv.tw/algorithm-complexity-theory.html)
#### Lattice
- Shortest Vector Problem
- Given basis of a vector spcace V and ...><
- 

#### Recall RSA

#### [NTRUEncryption](https://en.wikipedia.org/wiki/NTRUEncrypt)

#### [McEliece](https://en.wikipedia.org/wiki/McEliece_cryptosystem)

### Conclusion

### QA
- 國內商機(業界)
- 軍方需求
**Post-quantum cryptography(PQC)**
- 台灣現在還在開發量子電腦