[Thoughts]-On the interpretability conundrum

Of all the unknowable information, you can only get as much unknowable information out of a system as axioms you put in. Better understanding depends on the amount of input axioms. Better prediction requires better understanding. And better understanding makes a system more interpretable. Higher the interpretability of a system the higher we trust the system. Interpretability focuses on understanding the cause of the decision while inference focuses on the conclusion reached on the basis of evidence & reasoning(Causal & Bayesian).

However there’s a general notion of trade-off between the accuracy and interpretability of the system let alone the privacy aspect. The most affected domain by this trade-off is the medical domain. The problem is that the systems are heavily biased; be it racial, gender or any other biases that you can think of. The ML models can predict self-reported race even from corrupted, cropped, and noised medical images as opposed to medical experts. These ML models are making accurate decisions about racial classification using features that humans can’t even notice carefully, let alone analyse. (Banerjee et al., 2021)
In order to trust the system we need to break things down like in the first principles approach and then make a ground-up approach to interpretability/reasoning. Hence, I believe, mathematical methods like causal inference, differential & algebraic geometry, topology, stability theory, probabilistic methods, PDEs, information geometry & algorithmic information theory can help achieve better interpretability. I believe with mathematical proofs and truths we can achieve inductive/abductive reasoning and inference while discarding the role of medical domain experts because humans have a tendency to lie, mathematics doesn’t lie. Math was already there, humans just discovered it they didn’t invent it. Math should be the base of the prediction, inference and reasoning instead of a ‘Domain level expert human being’. Thus, in my opinion the model should learn from the mathematical proofs instead (learn from nature, don’t learn from humans).

'If I bet on humanity, I'd never cash a ticket. It didn't pay to trust another human being. They never had it from the beginning, whatever it took'. (Bukowski, 1978)