1. Bias in Training Data:
- Understand how to mitigate risks, address ethical concerns, and make informed decisions on model selection and fine-tuning.
2. Availability and Compatibility of Pre-Trained Models:
- Pre-trained models are available on repositories like TensorFlow Hub, PyTorch Hub, Hugging Face, etc.
- Check for:
- Compatibility with your framework, language, and environment.
- License and documentation.
- Regular updates and maintenance.
- Known issues or limitations.
3. Customization and Explainability:
- Modify or extend models (e.g., add new layers, classes, or features).
- Ensure the model is flexible, modular, and transparent.
- Look for models that provide tools to visualize or interpret their workings.
4. Interpretability vs. Explainability:
- Interpretability is explaining mathematically why a model makes predictions.
- Transparency = Interpretability (simple models can be interpreted easily).
- Foundation models are black boxes and not interpretable by design.
- Explainability: Approximate a complex model with a simpler, interpretable one.
- If interpretability is important, simpler models like linear regression or decision trees may be better.
5. Complexity of Models:
- Complex models can uncover intricate patterns but increase maintenance costs and reduce interpretability.
- Consider the costs, performance, and complexity trade-offs.
- Other factors: hardware constraints, maintenance updates, data privacy, and transfer learning.
