Neural Network Debugging Toolkit

Create a debugging toolkit for neural network training issues.

Common issues to diagnose:
- Training not converging
- Overfitting/underfitting
- Gradient problems
- Performance plateaus

Debugging toolkit components:
1. Gradient analysis:
   - Gradient flow visualization
   - Vanishing/exploding detection
   - Per-layer gradient statistics
2. Activation analysis:
   - Dead neuron detection
   - Activation distribution
   - Saturation monitoring
3. Loss landscape:
   - Loss surface visualization
   - Sharpness analysis
   - Local minima detection
4. Learning dynamics:
   - Learning rate finder
   - Batch size effects
   - Convergence analysis
5. Architecture debugging:
   - Layer connectivity
   - Parameter counting
   - Receptive field analysis
6. Data debugging:
   - Label noise detection
   - Class overlap
   - Data quality metrics
7. Training monitoring:
   - Real-time metrics
   - Early warning alerts
   - Checkpoint comparison

Provide actionable recommendations.