theory salon

« first day (747 days earlier) ← previous day next day → last day (2882 days later) »

12:00 AM

@vzn Step 1. Obtain a set of pictures of knots (2d representation)

@vzn Step 2 create a vector that represents the classes that is either -1 or 1 whether the knot is in the class or not in the class

@vzn Step3 Train the classifier with the input obtained from step 1 and step 2. Make sure that both vectors are of equal length

@vzn Step4 Draw knots and input them into the classifier and see if it gives you the correct output

@vzn Overfitting / VC dimension doesn't apply because if I learn the data or overfit in polynominal time then it doesn't really matter because all of the other unknotting algorithms are in NP or greater

vzn

12:53 AM

sounds novel/ intriguing. have attempted to apply ML myself to a number theory problem (collatz). have used genetic algorithms & weka clustering, & some "adaptive" heuristics that verge on ML.

can you descr #1/ #2 in more detail?

overfitting/ VC dimension are an aspect of virtually every ML problem...

@JoshuaHerman for #1 do you enumerate over built-in mathemtica knots? how many are there etc?

sounds like a really great open science prj & hope you write it up in more detail on web page somewhere or ideally (arxiv) paper...

Joshua Herman

@vzn katlas.math.toronto.edu/wiki/The_Rolfsen_Knot_Table is step #1

Joshua Herman

1:16 AM

@vzn Overfitting makes sense on numerical and string data but because my dataset is a table and also the vc dimension is exactly 2-3 because my input is images and my output is strings

@vzn Also I am trying to test the error rate of the dataset thats why i want other people to input data into it

@vzn And i have done LOOCV validation on one of the programs so I have tested for overfitting

vzn

@JoshuaHerman ok there are only a few dozen knots in the rolfsen table...?

there has to be some test for "generalization" otherwise the ML is just "memorizing" all the input data....?

think it would be cool if you could work this into a Theoretical Computer Science, Cross Validated or MathOverflow post somehow...

the fact that the jones polynomial is highly connected to this area does give some basic credence to ML "curve-fitting" approaches... however you may be one of the 1st in the entire world to try something like this... ML vs math problems is mostly uncharted territory/ terra incognita...

have long thought/ conjectured there may be substantial potential & maybe even breakhrus lurking around...

alas tho the "community" seems conservative/ hesitant & slow to experiment...

vzn

1:46 AM

btw my reason for asking about your cpu was that theano seems to be tuned for GPUs/ parallel processing.

Joshua Herman

@vzn I test generalization by inputing knots in that i draw.

@vzn the community doesn't have the time to learn all of these things

vzn

2:18 AM

@JoshuaHerman yeah, no time for "new technology" apparently! maybe they all use rotary phones also!

Joshua Herman

@vzn One of my professors had a macbook g4 until after i graduated

@vzn I had to teach him how to upgrade his computer......

@vzn for all of the private tutoring i got i owed him so like i was really ok with it and all

just because you have a phd in cs doesn't mean you know how to use computers

vzn

so you have a BS/ CS? what are you doing now? still in academia somehow?

Joshua Herman

and also that can be a good thing with the internet being extremely distracting

no I'm a web programmer i do this shit for fun

this is like my hobby

i don't do video games

anymore

vzn

yes over in Computer Science there is even some real opposition to real computers, sometimes approaching near comical proportions...

Joshua Herman

rather do machine learning

i had an interview with another company and they asked me what i do for a hobby and i responded mathematics

i think they thought i was joking when i am totally fucking serious

vzn

2:21 AM

Q: How can I teach computer science without using computers?

In some places in the world, people don't usually have access to (and hence little knowledge of) computers, and even if they have, hard- and software are outdated and usage plagued by power outages and such. Access to (good) books also tends to be lacking. How can I teach computer science under s...

education

@JoshuaHerman =D

in Computer Science, Jul 10 at 7:31, by Raphael

@StanShunpike CS is very little about computers and technology.

in Computer Science, Jul 10 at 7:31, by Raphael

If you want to learn how to work current technology, you won't find it there.

have you read about deep learning much? its really gaining traction/ "catching fire" lately

Joshua Herman

@vzn wolframcloud.com/objects/15e03097-a085-44f3-bc0a-9dfbf3c01083

vzn

think it is just a matter of time before a deep learning algorithm advances a math problem in a major way that humans have failed at. its already nearly happened with video game playing algorithms etc.

Joshua Herman

@vzn I have. I want to see what deep learning will do

vzn

its a bit stunning how little machine-learning items there are on Computer Science and Theoretical Computer Science...

Joshua Herman

well there are actually machine learning things on other stackexchange sites

vzn

2:27 AM

it is a crosscutting area that doesnt really fit too well on any (single) se it seems...

Joshua Herman

its like

do you know stats

and ai

and computers

welcome to machine learning!

what do you want to learn?

vzn

recently have been talking to a CS major from bangladesh re graph isomorphism & he is very emphatic about not learning code. he says its "not his thing" lol!

Joshua Herman

well everyone wants to learn big data! guess what most of the time you will overfit this shit

vzn

actually some cs theorists are finally starting to notice big data, eg suresh mentions it on his blog etc.... (but ofc never talks about it on Theoretical Computer Science) :(

Suresh Venkat ♦, Salt Lake City, UT

25k 4 67 238

ds.algorithms cc.complexity-theory soft-question reference-request cg.comp-geom big-list graph-theory

two mentions of ML here by Aaronson/ Venkatsubramanian!

More fun at the theory festival: plenary sessions – guest post by Sanjeev Arora / Arora, windowsontheory

@JoshuaHerman so did you write any papers/ prjs up for classes as an undergrad?

Joshua Herman

3:29 AM

@vzn well i tried to create a lambda calculus with a self distributive algebra

@vzn i got pretty far after undergrad with kauffman but we didn't know what to do with it after constructing it

@vzn Here is something you can play with wolframcloud.com/objects/15e03097-a085-44f3-bc0a-9dfbf3c01083

14 hours later…

vzn

5:21 PM

The New Laws of Explosive Networks / Quanta mag

> Researchers are uncovering the hidden laws that reveal how the Internet grows, how viruses spread, and how financial bubbles burst.

3 hours later…

vzn

8:10 PM

@ThomasKlimpel thx for the ref in your answer but... you might get more pts around here if you dont credit me :P

& think you have sort of proven my point, by answering your own question! maybe that is the end evolution of se restrictions in general! experts posting & answering their own questions!

hey, looking for a good overview/ survey of the connections between group theory & graph isomorphism! ideally online... fyi you might find this of interest...

in Room for Jim and vzn, Jun 30 at 18:12, by vzn

Group-Theoretic Algorithms and Graph Isomorphism / Hoffmann

...

Q: Testing a Computational Topology Algorithm

I have a algorithm in Computational Topology that I wish other computer scientists to test. Can I post it on the main site? I don't have that many contacts in computational topology.