1
$\begingroup$

Neural networks are the base for all(most?) of the machine learning / deep learning algorithms/programs.

Humans don't have fixed algorithms to decide or do something. Initially, we don't know how a snake is supposed to look. We learn that from experience. But our learning is never perfect. Once in a while, we mistake a black wire lying on the floor for a snake. Or we mistake the snake for a wire.

Given that neural networks are using the same technique as humans, are they inherently imperfect? Surely, we can make the self-driving network much better than best human, but that is because it got much more experience(24x7) than the human. But can we guarantee it's perfectness?(mathematically, or at least by testing, enough to make people trust in the programs)


I have this question because of the moral implications ML programs bring. I guess this will cause a lot of problem in adopting these technologies, especially when human life is at stake(eg. surgeon, self-driving cars, self-flying jets) as humans would rather rely on an imperfect human than a lesser imperfect AI.

$\endgroup$
1
$\begingroup$

It's better to think of neural networks as inspired by human biology, not that they're using exactly the same mechanism.

The notion of "inherently imperfect" isn't well-defined, so the question isn't meaningful. And, no, neural networks aren't always imperfect. For instance, suppose the learning task was "label everything 0". If you give a neural network a training set with one million examples, all of which have the label 0, then the neural network might well learn to output label 0 on all possible inputs.

As far as moral implications, it's not clear that it matters much why neural networks are imperfect; what's important is that they are often imperfect. But moral implications are probably more a matter of the study of ethics and philosophy than the science of computing, so I'll leave that there.

Finally, I encourage you to avoid binary thinking. Don't divide the world into "perfect" vs "imperfect". An engineer would say that in practice nothing is perfect; it's all about degrees of imperfection. One system might be good enough, or better than the reasonably available alternatives, even if it's not 100% perfect.

$\endgroup$

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.