NDBug
Neural Debugger
- NDBug is a non-invasive neural network add-on tool for profiling and debugging
- Comes with a suite of diagnostic metrics based on group and invariant theory
- Isolate and mark faulty filters for re-training, freezing the remainder network weights
- Identify redundant layers and neurons for network compression
- Operates with the core principle that accurracy increases with interpretability
Practical Uses
Development
Separate good architecture choices from bad during the initial network build
Deployment
Compress and optimize by isolating redundant blocks prior to production
Maintenance
Identify network retraining requirements. Isolate or add additional feature blocks marked for training