[ExI] alpha zero

[Dylan Distasio]

On Thu, Dec 7, 2017 at 1:29 PM, John Clark <johnkclark at gmail.com> wrote:

> On Thu, Dec 7, 2017 at 12:07 PM, Dylan Distasio <interzone at gmail.com>
> wrote:
>
>
>> ​> ​
>> This type of program still needs to be trained on a very specific problem,
>>
>
> ​It trained itself and it started with nothing but the basic rules and
> was able to beat the best in the world at it, human or machine, in one day
> .
>  And it didn't just do it with one problem it ​did it with 3 different
> ones, Chess being the least complex.
>

Again, reinforcement learning is very successful in certain scenarios,
particularly ones involving games.   This has been done with other things
like Super Mario Brothers, a popular video game where the system learns how
to play and what strategies work best to maximize a selected outcome.
These systems can be trained very quickly with enough hardware thrown at
training them, and the tech to train them scales very well in general if
architected properly.

I would still argue that this is very far from strong AI.



>
> ​> ​
>> there is no thought process going on behind it.
>>
>
> ​That is a strange statement. ​
>
> ​If you can teach yourself to be the best in the world at some complex
> task ​without "thought" then what's the point of "thought"? Who needs it?
>

It's not needed as I'm defining it (human level intelligence combined with
consciousness (whatever that is, but I think we're relatively good at
identifying it) for most species on the planet to thrive in their niches.
In fact, the odds of it evolving MAY be extremely low.  I will give you a
real world example of why these networks don't think, and why thought is
important.  I'm going to shift into image recognition for the example.
It is very easy to game these machine learning systems with an adversarial
attack that shifts pixel information that is essentially  undetectable to
the human eye but that will cause the system to misidentify a turtle as a
gun (for example).  These attacks BTW are likely to become a very large
problem as we rely more on machine learning behind infrastructure and
systems.  Once there is an incentive for these hacks, they will show up in
commercial places.

The image recognition is very accurate in general, but it is also very
brittle and subject to this type of gaming.  Thought would probably be
helpful in eliminating these false positives as it would allow to see
things in the larger context and ponder the most likely possibility based
on experience, and the ability to visualize scenarios ad hoc.  All the
system consists of is a series of probabilities output based on input
flowing across weights in a matrix.

These image recognition systems are variations on the deep learning
technology used here (minus the reinforcement learning).

The point of thought is to be able to generalize and make decisions with
sometimes very limited information based on experience and imagination.
This system is capable of nothing like that.   It is still very brittle
outside of the goal it has been trained on.  It would need to be retrained
for each new goal, and if you attempted to apply it to real life, you would
probably wind up with some very unexpected behaviors.


> ​> ​
>> It's quite startling at first glance to think that an end goal of
>> minimizing a loss function can generate so much razzle dazzle, but the math
>> behind these systems is actually not that complex.
>>
>
> ​But we know for a fact that the ​
> recipe for a mind
> ​ can't be very big, we must have that master learning algorithm so we can
> put a upper limit on it.​
> In the  human genome there are only 3 billion base pairs,
> ​ ​
> there are 4 bases so each base can represent 2 bits, there are 8 bits per
> byte so that comes out to 750 meg.   And all that 750 meg certainly can not
> ​ ​
> be used just for the master learning software algorithm, you've got to
> leave room for instructions on how to build a human body as well as the
> brain hardware.  So the information
> ​ ​must
>  contain wiring directions such as "wire up
> ​ ​
> a neuron this way and then repeat that
> ​ ​
> procedure exactly the same way
> ​ ​
> 42
> ​ ​
> billion times".
> ​ ​
> And the 750 meg isn't even efficiently coded, there is a ridiculous amount
> of redundancy in the human genome.
> ​ ​
> I would guess
> ​ ​
> the
> ​ ​
> master
> ​ ​
> learning algorithm is less than a meg in size, possibly a lot less.
>
> ​ John K Clark​
>

I don't think I disagree with you here, but I also think we're comparing
apples to oranges.  Deep learning neural nets appear to bear little
resemblance to how biological nervous systems actually work.
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.extropy.org/pipermail/extropy-chat/attachments/20171207/e6aecda7/attachment.html>