Skip to content

A random thought experiment on Infinite Monkey Theorem.

License

Notifications You must be signed in to change notification settings

yash-srivastava19/P-GLAm

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 

History

22 Commits
 
 
 
 
 
 
 
 
 
 
 
 
 
 

Repository files navigation

Inspiration and the Problem Statement

It was a random Sunday night, and I was wondering about language modelling when something clicked. There is this famous experiment in probability called the Infinite Monkey Theorem, which states that a monkey hitting keys at random on a typewriter keyboard for an infinite amount of time will almost surely type complete works of Shakespeare. In fact, the monkey would type every possible finite text an infinite number of time. The probability of happening such events is exteremly low but technically not zero. Let's use this fact together with the capabilities of Language Modelling we have today. We'll twist the rules a little bit to better asses the experiment.

Imagine yourself as the monkey in this scenario, but instead of having to type the entire Shakespeare, we'll be typing arithmetic equations. The domain is a little restricted, with little characters(0-9 and binary arithmetic operators) which is ideal given the scope of the experiment. We'll feed this arithmetic data to a GPT-2 inspired model, and check for the validity of the arithmetic equation. If this experiment is repeated for several iterations - What is the probability the random Language Model will spit a correct arithmetic equation(with no domain knowledge whatsoever) ?

How will be the experiment conducted ?

We'll be implementing GPT-2 from scratch - inspired by the excellent tutorial by @karparthy. But instead of providing some normal data, we'll be creating a dataset of arithmetic equations and feed it to the Language Model we'll be calling P-GLAm(Pascal GPT Language Arithmetic model - named after the founder of this idea, Blaise Pascal) from now on. Then, we'll allow P-GLAm yo spit out random arithmetic expressions and we'll check the validity of the statements using sympy evaluator and analyze the results. Can we really have valid arithmetic expressions being generated by P-GLAm ? We'll find out for sure.

Limits of the experiment.

I was greatly limited by the compute capability of hardware. I tried replicating the experiments on both Kaggle and Colaboratory(both Tesla T4 and P100 GPU) - but since I had access to free plan, I couldn't generate more than 5000 tokens without exhausting the resources.

Results of the experiment.

Will be updated after initial tests. But still, the results are promising, the P-GLaM baseline model(3.87 M parameters) 'improved' the probability of generating corrected mathematical expression from 0.54 to 0.59 - indicating little improvement. Any further conclusion needs more data.

About

A random thought experiment on Infinite Monkey Theorem.

Topics

Resources

License

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published

Languages