person Author: Process Fellows
The ML model is typically considered as a black box! This question sounds like it would be impossible? Let's see what kind of techniques could be considered and used:

• Pruning (weight thinning)
  • Idea: Removal of unimportant neurons or connections in the network (some weights have very little influence on the final result).
  • Various techniques (e.g. magnitude-based pruning) can be used to eliminate weights or entire neurons.
  • Test: Retraining or evaluation is used to check whether the prediction performance is maintained.
• Quantization (reduction of the precision of the parameters)
  • Idea: Reducing the precision of the model parameters (e.g. instead of 32-bit float values for weights, reduce to 16-bit or even 8-bit).
  • Modern hardware (e.g. TensorFlow Lite or special AI accelerators) efficiently supports such quantized models.
  • Test: The accuracy of the quantized model is measured in comparison to the original model.
• Knowledge distillation (knowledge transfer from large to small)
  • Idea: A large model (“teacher”) trains a smaller model (“student”).
  • A smaller model is trained to imitate the output of the larger model.
  • As a result, the smaller model can often make similar predictions, even if it has fewer parameters.
  • Test: Evaluation of the predictions of the “Student model” compared to the “Teacher model”.
  Tip: If TensorFlow or PyTorch is used, there are libraries such as TensorFlow Model Optimization or Torch Pruning that automate many of these techniques.