Deep Learning Notes
These are the patterns I reuse when moving from classical ML to deep learning workloads.1. Core Architectures
- CNNs: still relevant for document layout parsing + GlucosePro image analysis.
- RNNs/GRUs: used sparingly; prefer Transformers even for sequences.
- Transformers: default for language + multimodal experiments.
2. Training Stack
- Frameworks: PyTorch Lightning + Hugging Face Accelerate.
- Compute: RunPod + Lambda Labs for spot GPU; on-prem for sensitive data.
- Experiment tracking: Weights & Biases, MLflow for metadata.
3. Fine-Tuning Strategy
- Start with adapter/LoRA to reduce cost.
- Use domain-specific datasets (Thinki.sh prompts, productivity transcripts).
- Evaluate with targeted benchmarks (guardrail dataset, task-specific metrics).
- Distill into smaller models for edge deployment when needed.
4. Multimodal Experiments
- Image + text journaling for productivity dashboards.
- Audio + text for Nishabdham recitals (Whisper + GPT-4o audio).
5. Responsible AI Hooks
- Dataset audits, bias checks.
- Red-team prompts documented alongside releases.
- Human-in-the-loop review for new models touching finance/health.