gomader/swe3028

Capstone Design Project (Fall 2021)

SWE3028

Capstone Design Project (Fall 2021)

Team Building

• Team A [jjangdol]: YOON SEONGBIN (윤성빈), NAM DEUKYUN (남득윤), WEE SUNGEUN (위성은), and LEE DASOL (이다솔)
• Team B [bTeam]: Jisu Kim (김지수), Jinhwan Kim (김진환), Seyeon Park (박세연), and Mujin Gwak (곽무진)
• Team C [coturnix]: Seonghyun Ban (반성현), Dongyoung Choi(최동영), and Minseung Lee (이민승)
• Team D [EDITH]: Oinar Chingis (칭기즈), Kim Eunmin (김은민), Park Soohun (박수헌), and Gong He (공허)
• Team E [Exponential]: Pavlov Borislav Georgiev, KIM MINJAE (김민재), KIM YOUNGOH (김영오), and PARK GUERYANG (박거량)
• Team F [Fancy]: CHA MINJI (차민지), LEE EUNJI (이은지), JO DAEYEOL (조대열), and KIM DAEHEE (김대희)
• Team G [cookie&cream]: NAMKOONG BOMIN (남궁보민), KIM HANGYU (김한규), SUH JUWON (서주원), CHO GYEONGHYEON (조경현), and CHOI JAEHYUK (최재혁)
• Team H [The outsiders]: CHE SEUNGYUN (채승윤), UHM JIYONG (엄지용), LEE JISEOP (이지섭), JEONG CHAEWON (정채원), and HONG SEONGJUN (홍성준)

Team Proposals

• Team A [jjangdol]: GANdan-fontmaker: Web Service for Handwritten-Hangul Font Generation
• Team B [bTeam]: S-Gether: Web Application for Sharing Goal
• Team C [coturnix]: Explainable AI model for Stock Trading
• Team D [EDITH]: AI-Powered Anime Character Editing Web App
• Team E [Exponential]: Stock-loss Prevention: Mobile Application with CNN-LSTM Model for Predicting Sharp Rises and Falls in Stock Price
• Team F [Fancy]: Review Note Auto Generation Application
• Team G [cookie&cream]: Deep Learning Based Fashion Recommendation Application
• Team H [The outsiders]: CNN based Location Image Search and its Adaptation to Social Network

Common considerations for AI-based projects

• Supervised or unsupervised?
• Dataset you want to utilize for training?
  • Manual dataset preparation for labeling?
  • Number of dataset
  • Number of labels (supervised)
  • Ratio (training/validation/test)
• Goal (e.g., classifier)
  • Clarify the objective of the model
  • Narrow down your scope if too broad -> constraints
  • Set up the objective function (min/max)
• Algorithm or model
  • Reasoning that you choose that model?
  • CNN-based, RNN-based?, GAN, ensembles?
  • Difference between your model and prior work
  • Applying the existing model (as it is) might not work for your goal
• Input and output form
  • Input form to be fed into a model
  • e.g., words must be converted into vectors (like word2vec) in NLP
  • pre-processing of raw inputs
  • output: classified category (e.g., softmax) or others?
• Defining unique challenges for your own problem
• Existing techniques to take advantage of
  • Open sources
  • e.g., OCR (optical character recognition) for recognizing chars
• Limitations (every work has limitations)
  • Constraints
  • Out of scope
• Evaluation
  • Evaluation metrics? F1 (with FPR/FNR), accuracy or AUC? (supervised)
  • Assessment on GAN outputs?
  • Comparison with another model at least one comparison: performance comparison with previous approaches
    • Availability of comparing models (open source?)
• Computation resource (e.g., GPU)
  • Out of scope; you should find your own
  • Cloud GPU?