what i've been up to
An AI-generated summary of what I'm currently working on and learning.
Focus Areas
Deep Learning Fundamentals
Strong, math-grounded coverage of core neural network building blocks โ convolutions (2D, depthwise, pointwise), batch normalization, dropout, activation functions, and gradient computation. Posts derive formulas from first principles (e.g., why dL/dW = gยทxแต, why 3ร3 kernels dominate).
Edge AI & Model Deployment
Practical focus on making models run efficiently: INT8 quantization (PTQ, QAT), mixed-precision training with BF16, depthwise-separable convolutions for mobile, and parameter reduction techniques. Understanding the full pipeline from FP32 training to edge deployment.
Numerical Precision & Hardware
Deep understanding of floating point representation, precision-range tradeoffs, tensor core computation, and why BF16 works despite limited mantissa bits.
Linear Algebra for ML
PCA, ZCA whitening, eigenvalue decomposition, covariance matrices โ with clear connections to how these tools serve ML pipelines.
ML Theory
Bias-variance tradeoff, ensemble methods (bagging, boosting, GBM, XGBoost), regularization techniques.
Depth Indicators
| Topic | Depth |
|---|---|
| Convolutions (2D, depthwise, pointwise) | Deep |
| Quantization & mixed precision | Deep |
| Gradient derivations / backprop | Deep |
| Batch normalization | Moderate |
| Ensemble methods (RF, GBM, XGBoost) | Moderate |
| PCA / whitening | Moderate |
| Dropout / regularization | Moderate |
| LLM tooling (Claude usage) | Surface |
| Transformers / attention | Not yet covered |
| Generative models | Not yet covered |
Learning Trajectory
Progressing from foundational deep learning math โ deployment/optimization concerns โ knowledge systems. The arc suggests movement toward understanding the full ML lifecycle: theory โ implementation โ efficient deployment.
Technical Style
- Math-first: derives formulas before explaining intuition
- Comprehensive: posts are study-guide length, not quick notes
- Practical grounding: connects theory to real constraints (hardware, memory, latency)
- Sources content from AI discussions (ChatGPT, Claude) and synthesizes into structured write-ups