Re: [dev] What would a well-designed voice assistant program look like?

From: ssd <ssd_AT_mailless.org>
Date: Sat, 24 Mar 2018 15:25:18 +0100

Laslo,

* Laslo Hunhold 2018-03-24 13:00
> most of the deep learning folks don't know what the hell they're doing
> and it will cost us big time in the near or far future.
> Just throwing data at a problem you do not understand will get you
> better results but won't solve the fundamental issue of not
> understanding it.

While I share some of your concerns (even worse: without understanding
what data one throws, there is no chance to meaningfully interpret the
results), I know among deep learning folks there are knowledgeable
people, including e.g. some which converted to DL after decades of
developing other approaches to vision problems.

Back to the point, end-to-end approaches are well worth attention, even
though not modular. They seem to be able to find patterns in the coupled
sub-problems to solve the combined problem better or more efficiently
than if explicitly factorized and modeled separately.

cheers
--s
Received on Sat Mar 24 2018 - 15:25:18 CET