[ExI] Neural networks score higher than humans in reading and comprehension test
rafal.smigrodzki at gmail.com
Fri Jan 19 01:15:24 UTC 2018
On Tue, Jan 16, 2018 at 1:31 PM, Dylan Distasio <interzone at gmail.com> wrote:
> This is the Achilles heel of all the deep learning that I've been exposed
> to. Deep nets are also great at image recognition until an adversarial
> attack is injected that will fool a net, but not a human child. There is
> zero comprehension. The nets are dead behind the eyes. I believe the
> folks attempting to replicate actual biological neuron function (the bulk
> of deep learning is not much more than high school math and a gradient
> descent function to find a global minima) have the best chance of building
> something that actually understands what it is doing.
### I agree with the general tenor of your remarks but I would not go as
far as saying "zero comprehension". Current deep learning systems are very
brittle, much more so than humans but they are on to something - small
parts of the world models that humans build over decades of learning are
already present in the networks. They are not fleshed out with enough
cross-references to different but complementary representations of aspects
of the world, which is why they are brittle. I cannot guess whether the
Microsoft or Baidu systems have 1% or 10% of adult human understanding of
the world but I would argue it is not zero.
I also doubt that the secret of human intelligence is in human neurons.
Neurons have been around a long time and human neurons are about 99%
identical biologically to monkey neurons - but we have more of them and
especially more of them wired to perform high-level integrative data
analysis. In effect, our brain's networks are just deeper than other deep
learning networks in existence, whether animal or artificial, but still
using only high school math to do their job. I guess.
-------------- next part --------------
An HTML attachment was scrubbed...
More information about the extropy-chat