[ExI] A AI with a higher IQ than 75% of Americans

Rafal Smigrodzki rafal.smigrodzki at gmail.com
Mon Jan 30 00:49:38 UTC 2017


On Sun, Jan 29, 2017 at 1:13 PM, William Flynn Wallace <foozler83 at gmail.com>
wrote:

> The Raven matrix test is just one component of many.
>
> -Dave
>
> ​Yes, but it's a good one and one of the very few that is language-free
> and has high validity.
>
>
### Raven's matrices have been standardized on humans, obviously, and
humans are a peculiar type of embodied neural network system with a lot of
hard-wired modules that enable real-world functioning. An IQ test takes a
lot for granted - ability to walk, hold a pencil, fill out the test as a
mechanical challenge, find the person you need to hand it to, etc. As a
result, Raven's matrices are not a very good measure of the intelligence of
disembodied fragments of neural networks that are now the subject of deep
learning research projects. Such fragments replicate some small areas of a
human mind but they do not form an integrated whole capable of functioning
in the real world, and their ability to perform some elements of an IQ test
is not predictive of real world performance.

Our fascination with IQ testing stems from the test's ability to predict
real-world function, including the eternally important who-whom question,
and not just silly puzzle solving. Deep learning network IQ testing is
therefore for now less interesting but once the separate deep learning
modules become integrated into human-like systems with real-world
performance, their IQ will be acutely interesting in the who-whom context.

This said, it is amazing that tiny deep learning neural networks with mere
billions of parameters manage to equal the performance of middling-sized
chunks of brain (e.g. the subcortical and cortical visual processing
centers).

Rafał
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.extropy.org/pipermail/extropy-chat/attachments/20170129/67ea6e18/attachment.html>


More information about the extropy-chat mailing list