Log In
Sign Up
Jobs
Job Search
Explore all available job openings across industries and locations.
Company Search
Find your dream jobs categorized by company names.
Themed Jobs
Discover job opportunities organized by specific themes or industries.
Download our App
Tools
Resume
Create your job-winning resume using our free resume builder.
Portfolio
Showcase your skills and projects with a professional portfolio.
Resume
Create your job-winning resume using our free resume builder.
Resume Builder
Make a resume for free.
Resume Templates
Access our extensive library of professional & ready-to-use templates.
Resume Examples
Get inspired by real resume examples to create your own.
Occupation Guide
Access resume writing guides tailored for different professions.
Resume Help
Get expert advice on all things resume from our team of recruitment specialists.
Portfolio
Showcase your skills and projects with a professional portfolio.
Portfolio Maker
Create a professional portfolio to highlight your skills and projects.
Portfolio Gallery
Browse through our collection of real portfolios for inspiration and networking.
Resources
Articles
Read insightful articles on career development, job search strategies, and more.
View All Articles
Job Search Guide
Resume & CV
Cover Letter
Portfolio
Interview Skills
Job Search Tips
Industry & Job Overview
Career Guidance
Career Planning
Career Tools
Career Development
Personal Branding
Success Stories
Success Stories
Business Excellence
People Operations
Recruitment & HR
About CakeResume
People & Culture
News & Updates
Events
Featured Reads
Resume & CV
What to Write in an Email When Sending a Resume [+ Examples & Tips]
Read More
Hire
Talent Search
Find Resumes.
Job Posting
Start for Free.
Recruitment Service
Acquire Talent.
Employer of Record (EOR)
Empower Your Business in Taiwan.
Employer Branding
Build and promote your employer brand.
Pricing
Job Posting Plans
Talent Search Plans
Resume Builder Plans
Build your Network
My Network
Access your personal network connections and manage your contacts.
CakeResume Meet
Expand your professional network by meeting and connecting with other users.
Community
Engage with other users through discussions, forums, and networking events.
Download our App
Build your Network
My Network
Access your personal network connections and manage your contacts.
CakeResume Meet
Expand your professional network by meeting and connecting with other users.
Community
Engage with other users through discussions, forums, and networking events.
Log In
Sign Up
Portfolios
趙梓吟
Zero-Shot Learning for Novel Attribut...
Zero-Shot Learning for Novel Attribute Synthesize
Zero-Shot Learning for Novel Attribute Synthesize
By
趙梓吟
國立交通大學研究助理
・
Hsinchu City, Taiwan 300
Connect
Training Phase
Infereence Phase
[負責項目] 演算法之設計與實踐(論文為共同第一作者) [專案目標] 為資料集標註細緻的屬性標籤(attribute label)所需的人力成本很龐大。以CUB鳥類辨識資料為例,每張照片需要標註312個屬性標籤,諸如"有紅翅膀"、 "有綠色胸"……等。因此,本案目標為開發可有條件的減輕標註負擔的技術。 人類可以理解語意間的組合,並且透過邏輯組合舉一反三的運用。例如,教一個人辨識紅翅膀、藍翅膀後,他很大機率就能理解"翅膀"這個概念,也能和其他已知概念例如"藍色"組合運用。受此啟發,我們開發一種可對訓練完成之影像特徵識別器進行邏輯操作的網路。例如,給定訓練好的"藍胸"、"綠胸"、"紅翅膀"、"紅頭"特徵識別器權重,可以透過邏輯運算模型以(紅頭⋂紅翅膀)得到"紅", (藍胸⋂綠胸)得到"胸",最後再利用(翅紅∪胸),在沒有任何"紅胸"訓練影像的情況下得到"紅胸"的識別器。 如此一來,若將來某新資料集需要標註屬性標籤,且其屬性標籤間能夠用邏輯組合互相描述,便只需人工標註最低限度種類,剩餘可由本方法自動標註完成。 [專案績效] 1. 本研究展示於標註CUB dataset上,人工標註32種attribute標籤後,可透過本方法自動標註出另外207種新穎類型的標籤。 2. 使用本方法標記出的attribute標籤作為類別語意(class semantic)訓練不同Generalized zero-shot classification的演算法,可以達到與使用人類標註attribute標籤作為類別語意訓練相近(甚至略佳)之效能。 [專案進度] 完成,研究成果已經投遞至研討會進行同行審議[1] [專案挑戰] 1. 克服人類與機器學習模型的語意間隙(Semantic gap),訓練之特徵偵測器(attribute detector)需能運用與人類理解一致的特性執行判斷。(例如,判斷一張圖片中含有”紅色翅膀”,則其響應位置必須落在紅色翅膀上) 2. 如何訓練可以對特徵偵測器權重執行邏輯運算(聯集/交集)的模型,確保組合出之新特徵偵測器能正確捕捉對應特徵。 [專案設計] 為應對上述挑戰,本案的核心想法如下: 第一個步驟為根據現有人工屬性標籤種類訓練出對應屬性辨識器(即Seen attribute detector)。對應挑戰1,由於訓練出來的辨識器將會被用來做邏輯運算,因此確保其響應的特徵正確至關重要。我們做出兩個設計: 1. 辨識器的權重皆為正值:藉由這個設計可以強迫辨識器學習尋找正面證據來做辨識。若把圖片特徵向量中的每個維度數值理解為"含有有多少這種特徵的成分",那麼描述一種屬性的辨識器就是利用"成分組成比例"的方式來呈現。如果要準確描述一種屬性,例如"紅色翅膀",我們希望辨識器是因為圖片"紅色"比例高且"翅膀"比例高而判斷其"高機率含有紅色翅膀",而非因為其他成分,例如"藍色"比例高就降低判斷"含有紅色翅膀"的機率。 2. 利用損失函數約束辨識器的響應盡量集中(L_umc):藉由這個設計可以避免辨識器學習過於概括、模糊的特徵。 接著,我們利用剛剛訓練好的Seen attribute detector的權重作為訓練資料,進一步訓練intersection網路(Union在我們的實驗中發現使用簡單的average效果較佳)。具體來說,intersection和Union的輸入都是兩個偵測器權重,輸出則為邏輯組合後的對應語意偵測器權重。因此我們利用手邊擁有的Seen attribute detector以邏輯組合互相描述,再重構損失(L_rec)來訓練網路。 最後,待網路訓練完畢,即可使用這兩個邏輯操作組合出新種類的屬性偵測器(Unseen attribute detector),並使用這些偵測器標註資料集。 [1] Y. -H. Li*, T. -Y. Chao*, C. -C. Huang, P. -Y. Chen, W. -C. Chiu, “Make an Omelette with Breaking Eggs: Zero-Shot Learning for Novel Attribute Synthesis”, arXiv:2111.14182 (Preprint), 2021. (* equal contribution.)
Report
Please login to comment.
Post
趙梓吟
Connect
Published: May 30th 2022
49
8
0
Tools
Python
Representation learning
Deep learning
Zero-shot learning
Computer vision
Share
Other works from
趙梓吟
建築數聚時代
智慧建築+
低標記成本的影像式停車場空位偵測模型訓練方法