[ExI] One Per Cent: Watson, the supercomputer genius, heads for the cloud

[BillK]

On Sat, Sep 22, 2012 at 12:03 PM, Anders Sandberg wrote:
> The problem is nonhuman biases. This kind of complex system makes certain
> implicit assumptions that can be very subtle, yet affect what it does. If
> there is an obvious performance metric to check that it does better than
> humans it doesn't matter much. But imagine that after a few years of use it
> was discovered that it made more reliable diagnoses for members of majority
> ethnic groups (more data available), or for some reason tended to
> underdiagnose certain diseases in certain people according to a complex
> pattern. There is nothing directly programmed in to achieve this, it is just
> an emergent effect of the machine learning algorithm and the existing data.
>

Watson will also make mistakes. It did not answer the Jeopardy
questions perfectly. Just on average much better than the best humans.

Hopefully, in all the data analysis fields that Watson is applied to a
committee of humans will apply a 'reasonableness' test to the answers.
I think that when Watson makes a mistake it will be a really
ridiculous mistake that should be easily spotted. As in the Jeopardy
mistakes it made.

Subtle bias such as you suggest would be more difficult to detect (if
it occurs). But then doctors have applied wrong treatments and
inadequate treatments for years - even generations.  We won't get
perfection, but Watson should still be on average better than human
doctors.

>
> I think that has been the case for a long time. As soon as either the
> mechanism of action is outside normal experience (e.g. electrical devices)
> or the system becomes nontrivial (e.g. a clockwork) most people just give up
> and accept that it works without any attempt at understanding. Most
> Victorians could likely not explain a steam engine.
>
> People who peek into black boxes have always been a minority. Maybe we would
> be better off to ensure that there are more of them (for reasons of
> reliability, democracy, innovation), but in terms of reaping the benefits of
> technology we probably don't need many.
>

More engineers. That's what we need!  :)
I think modern civilisation is rather more extreme than the
Victorians. We seem to be pushing ourselves into the situation that if
our gadgets fail, we become helpless blobs wailing in despair.


BillK