HRyu
← Notes

AI & Technology

Machine Learning

Covers: GAN project, ML domain survey, time-series methods, common debug patterns

Machine Learning

Covers: GAN project, ML domain survey, time-series methods, common debug patterns. See also [[cybersecurity-thesis]] and [[ai-in-industry]].

GAN Project — Synthetic Financial Data Generation

Source: raw-sources/AI Projects/GAN project on Google Colab.md

Goal: Generate synthetic OHLCV + indicator data that statistically matches real market data (AAPL).

Pipeline

1. Data Download

yfinance.download("AAPL", start="2015-01-01", end="2024-01-01")

2. Feature Engineering

  • Moving averages: MA10, MA20, MA50, MA80, MA100, MA120
  • Daily return: (Close - Open) / Open

3. Preprocessing

  • MinMaxScaler — scale all features to [0, 1]

4. GAN Architecture

Generator:

Dense(128) → LeakyReLU → Dense(256) → LeakyReLU → Dense(512) → LeakyReLU → Dense(n_features)

Discriminator:

Dense(512) → LeakyReLU → Dropout(0.3) → Dense(256) → LeakyReLU → Dropout(0.3) → Dense(128) → LeakyReLU → Dropout(0.3) → Dense(1, sigmoid)

5. Training

  • 10,000 epochs, batch size 64
  • Adam optimizer, lr = 0.00001
  • Alternating generator/discriminator updates

6. Generation & Validation

  • Generate 100,000 synthetic samples
  • KL divergence per feature (real vs synthetic)
  • PCA visualization for distributional similarity

7. LSTM Variant

  • Sequence modeling: (batch, timesteps, features) → LSTM → Dense
  • Captures temporal dependencies not captured by dense GAN

Common Bugs

  • Prophet: ds/y column naming; timezone conflicts
  • LSTM: shape mismatch (3D tensor: batch × timestep × features)
  • ARIMA: index alignment errors after resampling
  • General: NaN propagation through feature engineering chain; column creation order bugs

ML Domain Survey

Source: raw-sources/AI Projects/List of Different ML Domains.md

8 Key Domains

Domain Research Frontiers
Reinforcement Learning Robotics dexterity, RLHF for alignment, plasma control (nuclear fusion)
NLP Hallucination reduction, interpretability, efficiency (smaller models), low-resource languages
Computer Vision Object detection, semantic segmentation, video understanding, depth estimation, point clouds
Audio Processing ASR at scale, generative audio/music, real-time translation
Multimodal Text-to-image alignment, cross-modal retrieval
Graph Neural Networks Scaling to large graphs, generalization across graph types
Applied AI Industry-specific deployment (healthcare, finance, logistics)
Evolutionary / Meta Learning Learning to learn, few-shot generalization

Hakyun's Active ML Stack

Task Tools
Time series forecasting Prophet, ARIMA, SARIMAX, pmdarima (auto_arima)
Deep learning TensorFlow/Keras (LSTM), XGBoost (basic)
Classical ML scikit-learn
Visualization Plotly, Matplotlib
Scaling MinMaxScaler, RobustScaler
Splitting Time-based train/test splits (no data leakage)

Common Workflows

  • 30-day forward predictions
  • Feature engineering: moving averages (MA), RSI, MACD, Bollinger Bands, OBV, ATR, VWAP
  • Time-based splits (no random shuffling — preserves temporal integrity)

Claude Code Skills (AI Tooling)

Source: raw-sources/AI Projects/Skills.md

Four plugin recommendations for Claude Code:

  1. superpowers-lab — experimental features + token reduction
  2. claude-mem — persistent memory across sessions (superseded; use Claude's native auto-memory system instead — see [[claude-code-tools]])
  3. context-manager — filter relevant context to reduce noise
  4. claude-context — smart search over codebase

Agent Identity Philosophy

Source: raw-sources/AI Projects/Research & ML/MD files - What they are.md

OpenClaw's four primitives:

  1. Persistent identity (who the agent is across sessions)
  2. Periodic autonomy (scheduled tasks without user input)
  3. Accumulated memory (what it has learned over time)
  4. Social context (who it's talking to, relationship history)

On Claude's Soul Document: Hakyun treats Claude's Constitutional AI as a founding civilizational text — analogous to Harari's intersubjective realities, Durkheim's collective conscience. Five unresolved problems: Legitimacy, Plurality, Drift, Mythology of Creator, Observer Effect.

Related Pages

[[cybersecurity-thesis]] | [[ai-in-industry]] | [[active-projects]] | [[hakyun-ryu]] | [[claude-code-tools]]