🧠 Daily Study Log [2025-10-10]
Generated a new self-reflective idea, studied the InternVideo2 (Day 3) pretraining setup, and started a new economic sentiment analysis project comparing media tone with real market indicators.

💡 Idea

• Proposed Mirror Planner – Visual Reflection Meets Productivity.
• Inspired by the habit of writing on a whiteboard and self-reflection in front of a mirror.
• Integrates the SAVERS routine’s A (Affirmation) and V (Visualization) components into a mirror-based planning tool, encouraging both goal-setting and mindful self-checking.

📖 Paper Study – InternVideo2 (CVPR 2024, Day 3)

• Studied Pretraining Setup & Datasets section of InternVideo2: Scaling Video Foundation Models for Multimodal Understanding.
• Key Insights:
  • Scaled pretraining on InternVid-400M, one of the largest video–text datasets.
  • Three-stage pretraining pipeline:
    1. Masked Video Modeling (MVM): spatial-temporal representation learning.
    2. Multimodal Contrastive Learning (CL): alignment across video, text, audio, and speech.
    3. Next-Token Prediction (GEN): enables generative multimodal reasoning.
  • Fine-tuning Tasks: video-text retrieval, action recognition, captioning, and AVQA.
• Findings:
  • Massive and diverse data greatly improves zero-shot generalization.
  • Training uses DeepSpeed, FlashAttention, and FP16/BF16 precision for efficiency.
• Takeaway:
  • Multimodal scaling and progressive curriculum are key to robust video foundation models.

💻 Personal Project – Social Sentiment Monitor

• Initiated new project: Social Sentiment Monitor – Economic Sentiment vs. Market Reality.
• Goal: Analyze emotional tone of economic news and compare with market indices (KODEX200, KOSPI).
• Pipeline Overview:
  1. Crawl Naver economic news (keywords: economy, inflation, interest rate, etc.).
  2. Apply KoBERT / KcELECTRA for sentiment classification.
  3. Collect market data from Yahoo Finance (yfinance).
  4. Compare sentiment trends with market movements through visualization & correlation.
• Expected Outcome:
  • Determine if negative sentiment precedes market downturns.
  • Identify outlet-specific tone bias.
  • Explore causal patterns (e.g., Granger causality).
• Next Step:
  • Complete news crawling & sentiment labeling within Week 1 and Week 2 milestones.

✅ TL;DR

📍 Proposed “Mirror Planner” – reflection-based productivity idea
📍 Studied InternVideo2 (Day 3) – multimodal pretraining & datasets
📍 Launched “Social Sentiment Monitor” – linking economic sentiment with real markets