On Mon, Feb 10, 2020, 11:48 AM stefan.reich.maker.of.eye via AGI <
[email protected]> wrote:

> Matt, I'd like to take you by your word. You said an AI an be made out of
> a compressor. Show me an example of an AI solving any task built using that
> method. I don't think I've come across any one yet.
>

No compressor can pass the Turing test yet. But I have used compressors for
anomaly detection and machine learning. For anomaly detection you have some
training data X and test data Y and you want to measure whether Y resembles
the samples in X or is anomalous. For normal Y, you have high mutual
information and the compressed size together, C(XY), is close to C(X). If Y
is unexpected, C(XY) is close to compressing separately C(X) + C(Y). I used
this method with PAQ in some postdoc work in time series anomaly detection
to detect faulty valves for the space shuttle program.

Years later there was some discussion of MOSES on the OpenCog list. MOSES
is an evolutionary learner that takes training data and finds a program
that reproduces it to predict future data. In one test, it took 30 minutes
to solve the 4 bit parity problem, which outputs 1 if an odd number of
input bits are set.

I wanted to see if a compressor could solve this. I prepared some training
data using one byte per instance. In each byte the 4 input bits were picked
at random, one output bit was set appropriately, and the other 3 bits were
not used and set to 0. In order to compress to less than 5 bits per
character (ideally 4), the compressor has to learn the function. gzip does
this in about 1 ms.


------------------------------------------
Artificial General Intelligence List: AGI
Permalink: 
https://agi.topicbox.com/groups/agi/Tb12281a4ff6a2b10-Mfcad84b53e84e83970262d5a
Delivery options: https://agi.topicbox.com/groups/agi/subscription

Reply via email to