[mannykannot]

If you look at the specific issue that provoked this law, you will see issues that can - and should - be regulated. the city was using DNA matching software that turns out to have demonstrable errors, yet the office of the medical examiner stonewalled all attempts to have it audited until, thankfully, the courts forced the issue.

This sort of unjustifiable secrecy (the accused absolutely have a right to examine the premises of the accusation) can be regulated. Unfortunately, this law substitutes nebulous criteria which, no matter how worthy, are likely to turn a clear-cut situation into a tar-pit of legal wrangling that the victims cannot afford to enter.

The chief medical examiner is still holding fast on the very dubious claim that these flaws raise no doubts about the convictions in other cases where it was used, another area where I think specific legislation is needed.

> How do you regulate algorithmic learning when we don't fully understand how learning works?

This was not one of those cases. There are, however, cases - and this would be one if it applied - where it is reasonable to say that you can't use it until you can explain how it works.

[tomjen3]

In this particular context, maybe, but in general - who cares if you can't explain all the values in the neural network, when it is demonstrately safer than a human driver.

[Daishiman]

You can't say "demonstrably" until you can specify the full breadth of the algorithm's specs. Something being "demonstrable" is something incredibly hard to achieve in certain classes of NNs.

If it breaks on edge cases, that's important.