tl;dr: 
 * I think there should be more attention given to Juyang Weng's more recent 
(2008 - present) papers on the Developmental Networks as a source of solutions 
to creating AGI (WWN-1 to WWN-9) 
<http://www.cse.msu.edu/~weng/research/LM.html#:~:text=an%20account%20of%20wwn-1%20to%20wwn-9%3A>.
 
 * Also, I've run into some problems with the models he described and I've been 
unable to fix them. The solutions I've tried just don't work well and would 
like to get some ideas as to why they didn't work.

A few years back I went on a real AI kick where I wanted to learn as much as I 
could and see if I could understand why AGI was never achieved. When I was 
younger I used to browse AI forums and it always seemed like it was so close.

So I spent a few years re-learning math concepts I struggled with: statistics, 
linear algebra, calculus, etc.. and reading papers/books on a wide array of 
techniques from deep learning, cognitive architectures, automated planning, 
symbolic regression, and more. Not that I'm particularly knowledgeable now, but 
I understand the situation much better than I did when I was younger.

My day job is a software engineer and I've been doing it for a long time, so I 
leveraged that to test models that I've found in those papers by reproducing 
them. Some of which taking a very long time to do, since I'm not an academic 
and reading papers is a pain.

While attempting to reproduce Weng's work I ran into the following problem. Or 
I'd rather say, things holding back the model from having the potential to be 
used to make something resembling an AGI.

 * Equivariance - This is a common issue with machine learning models I am 
aware. But at least for this model it seems like something that could be 
overcome if a good representation could be found. Weng seemed to dismiss these 
problems in his work by explaining that with enough data these could be 
effectively achieved. Which seems like a non-answer as it would negate much of 
reason to use a model like this, which primarily is efficiency and generality.
   * Location Equivariance
   * Scale Equivariance
   * Rotation Equivariance

My best attempt at dealing with equivariance is as follows and I unfortunately 
don't understand why these solutions don't work. I believe this is a failing on 
my part to understand the math or how these things interact with each other.
 * For scale, rotation, and local location equivariance I have tried to 
represent the input data using Local Moran's I (LISA) or Geary's C. For 
instance with an input image, the LISA would be calculated using a queen 
contiguity weights matrix. Then once normalized, something like the Euclidean 
distance, cosine similarity, or the correlation coefficient can be used for a 
similarity measure. Typically this is done for some kind of classification 
task. It seems to work ok (as it identifies classes mostly correct) in simple 
tests but not in more complicated ones.
 * Using the energy distance instead of other distance metrics for judging 
similarity. I wanted to do this to deal with global location invariance where 
something may be the same locally but scrambled around globally. The results 
for this are not good (for anything other than that specific problem), 
especially when compared to just using any of the previously mentioned 
similarity measures.
 * When representing data using LISA and averaging (a better word for this 
escapes me) it, I am not sure whether it can reliably be used for similarity 
measures. This isn't really as much of a problem when there is an assumption 
that a particular class of data will not have large fluctuations in any 
dimension (the aforementioned position, rotation, or scale for example).
I'd appreciate any tips, suggestions, or criticism, thanks!
------------------------------------------
Artificial General Intelligence List: AGI
Permalink: 
https://agi.topicbox.com/groups/agi/Tabd1763b276303cf-M3a22afda2c0a08ddf99aeeba
Delivery options: https://agi.topicbox.com/groups/agi/subscription

Reply via email to