2:10 AM
Hey vzn

2:34 AM
@JoshuaHerman hey whats up

2:55 AM
@vzn I'm working on a machine learning project you interested?

3:15 AM
@JoshuaHerman sure tell me about it! :)
(have huge pile of AI links waiting to be written up...)

3:32 AM
@vzn So i'm trying to compute knot invarants by reducing the problem to a computer vision problem which basically requires a precomputed labeled training set.
I am using random forests as my model
and i use knot tables as my training set
so i basically tell it. here are a set of chiral and amphichiral knots and then see if can generalize that information into knots that i haven't programed

3:53 AM
so it computes not only the training set but derivations of the training set

@JoshuaHerman very cool! could it be learning some variant of the jones polynomial? wondering, what math problem does the classification reduce to?
which ML pkg are you using?
strange coincidence, google finds a question on knot theory vs ML from few days ago on Theoretical Computer Science but it was deleted by author... not you? didnt see it myself...
there is this:
have been musing some lately that maybe new Deep Learning field might be applied to math/ cs problems soon with some results!
do you know of any other related refs?
...
sigh, graph isomorphism problem killed in just a few hrs... :(
Alex A., Seattle, WA
5.8k 2 12 39
13.3k 1 20 83
hildred, near mountians
773 4 26

4 hours later…
8:34 AM
I think the question got closed because it was too broad, not because it was related to graph isomorphism. Part of the art of problem solving is finding sufficiently small subproblems which can be solved in reasonable time and space. For example, I asked:
5

Is the finite inverse semigroup isomorphism problem GI-complete? Here the finite inverse semigroups are assumed to be given by their multiplication tables.

This might seem like a reasonably short question. However, it implicitly contains the two subquestions "Is group isomorphism GI-complete" and "Is (semi-)lattice isomorphism GI-complete". The first subquestion is a well known open problem, and the second question contains the subquestion "Is poset isomorphism GI-complete".
Now bipartite graph isomorphism is known to be GI-complete, and is basically equivalent to poset isomorphism. And poset isomorphism can be reduced to (semi-)lattice isomorphism by via the Dedekind-McNeil completion. Oh wait, the size of the Dedekind-McNeil completion can be exponentially bigger than the size of the original poset. And I already thought I had finally solved this question :-(

5 hours later…
1:38 PM
@vzn I don't think it's learning an invariant I think it's doing something else

1 hour later…
3:02 PM
@ThomasKlimpel hi sometimes think se taken to extremes boils down to a catch 22. have long thought the "broad" criteria doesnt make a lot of sense as often applied. questions are closed as "too broad" which nobody answers. seems to me the broad criteria should be leveled against questions that get too many answers & that its unfair to apply it a priori. also comparing `cstheory` with `codegolf` is quite a stretch.
think a lot of the justifications about closing questions are hindsight rationalizations.

2 hours later…
5:07 PM
@vzn I tried clustering the data and I am trying to figure out what the clusters mean.....
The clusters could be a new knot invariant... The problem is is that I don't know what criteria it is doing to make the cluster
@vzn Ihabe seen that google result and also I searched arxiv for other results and found nothing.
@vzn I'm using Mathematica but I am going to move to Python probably scikit learn and also theano

@vzn Well, I often tend to ask broader questions than necessary. I notice this most badly when I try to answer my own questions. I guess if I had just asked "Is (semi-)lattice isomorphism GI-complete", somebody would have answered "yes" and cited a paper with the proof.
But because I didn't, I probably have to find that paper with the proof myself. Going through the references listed in the wikipedia article will probably be sufficient for this, but this means quite some work for me, just because I didn't narrow down my question sufficiently. OK, I'll do it now. Let's see how many of these references can be accessed directly via google...

5:37 PM
@JoshuaHerman are you using a ML library? can you be more specific about the ML algorithm? are you running on 1 cpu? yes interpretation/ analysis of the models can be very challenging in ML in general...
think its semi random what questions succeed/ fail, se values them low/ randomly in general, somewhat like ... sperm!

3 hours later…
9:06 PM
@vzn Mathematica has a built it ml library I'm using random forests to classify everything

9:33 PM
@JoshuaHerman ok would be great to hear more maybe you could write it up on your home pg or something.... are you collaborating with anyone? intend to write it as a paper?

9:49 PM
I demoed it to Louis Kauffman but he only really verified that it was working

10:37 PM
@vzn I don't really have anyone to collaborate with.
@vzn I have my brother replicating my results.
@vzn Model Name: MacBook Pro
Model Identifier: MacBookPro12,1
Processor Name: Intel Core i5
Processor Speed: 2.7 GHz
Number of Processors: 1
Total Number of Cores: 2
L2 Cache (per Core): 256 KB
L3 Cache: 3 MB
Memory: 8 GB
Boot ROM Version: MBP121.0167.B07
SMC Version (system): 2.28f7
Serial Number (system): C02P8A7MFVH5
Hardware UUID: 863E8E68-B7B3-530F-A79A-7FBECDF9BE19
@vzn Also this model doesn't have a GPU :)
@vzn Or mathematica can't really access it its integrated graphics.

10:55 PM
do you have any plans to write up the results? where did you get the data?

11:18 PM
@vzn The data is built into mathematica
@vzn i'm trying to write up the results

11:36 PM
@vzn I use the class of the knots to create a binary function either -1 or 1 if the knot is in the set or not

11:59 PM
@vzn Here are my methods