/sci/ - Science & Math » Thread #12533287

109KiB, 424x550, 1606330316293.jpg

View Same Google iqdb SauceNAO

AI and laziness

Anonymous Fri 01 Jan 20:29:49 2021 No.12533287 View Reply Original Report

Quoted By: >>12533615 >>12534346 >>12534391 >>12534422 >>12535535

I understand that Machine Learning is a necessary method in order to solve problems that we otherwise have no means yet to handle.

But doesn't that mean researchers will never have the incentive to sit down and try to find a logical, structured solution to the problem?
All we have is an algorithm that WE don't understand, only the machine. So the solution has been found, but the knowledge generated is questionable.

Anonymous

Anonymous Fri 01 Jan 2021 22:24:43 No.12533615 Report

Quoted By:

>>12533287
Good machine learning research does improve general understanding of a domain.

In natural language processing, there are theoretic formulations for a generative probabilistic model that could perfectly describe any natural language (by modeling the conditional probability of any word being generated after an arbitrarily long sequence of words). However, such a formulation includes more conditional probabilities than there are atoms in the universe. What ends up happening is linguists find imperfect approximations to the true generative model.

Machine learning research improves understanding by highlighting which approximations are more “valid”. In early works, linguists tried constructing syntax trees to compute language models, but such solutions were outperformed by learning-based solutions such as LDA. The success of LDA reveals weaknesses with earlier language models such as syntax parsing ambiguities. Learning covers up these ambiguities by deriving parameters that model ambiguities using data. Doing something like this by hand is intractable (it would be a waste of time finding the probability of every word in the dictionary given every previous word). In a similar vein, recent progress such as BERT can model even more ambiguities that LDA (ex. removing Markov assumption), which further improves performance. I think the key problem here is some problems cannot be modeled by tractable formulas such as F=ma, thus these problems require a more “statistical” approach.

I do agree that the trend of making models relying on hundreds of GPUs number crunching is bullshit, but interpretability is still a priority for many researchers in the field.

Anonymous

View Same Google iqdb SauceNAO Screenshot_20210101-192031_Adobe (...).jpg, 1MiB, 1080x2220

Anonymous Sat 02 Jan 2021 02:21:29 No.12534346 Report

Quoted By:

>>12533287
good question

problems belong to something called complexity classes. so problems are not as unique as they seem. many problems across different fields are actually the same problem in terms of complexity. so its not desirable to solve most problems uniquely. separating complexity classes and improving the best algorithms fot each one is currently (mostly) beyond the capability of machine learning.

most researches dont want to solve unique problems in many cases, rather find a solution for an analagous problem and adapt it to their data.

efficiency of algorithms typically tends toward an ideal, theoretical running time asymptotically. therefore researches are forever finding slightly better methods but usually not the perfect one.

for example, problems in networking computers together over the internet may use identical techniques and algorithms to protein folding in biology, due to the problems being quite literally the same problem.

pic related is how computational resources such as time and space can be formalized for adaptation to any problem.

Anonymous

Anonymous Sat 02 Jan 2021 02:53:19 No.12534391 Report

Quoted By:

>>12533287
You're more or less correct - Machine Learning has become a lazy, brute-forcing method for getting around problems that would otherwise require creative problem-solving and it pisses me off to no end because there's no guarantee that ML results are actually correct, just consistent with what the AI was already trained on.

Saw a great example of this at an interdisciplinary conference a few years ago. Someone presented an example of using ML to reconstruct an image of a fish from a low-resolution version with compression artifacts. However it only worked because the presenter had spent all this time training the ML to recognize images of fish and then reconstruct shitty pixelated images of a fish. If you fed any other image into the same program you got some shitty distorted picture back where the AI tried to force fishy features in everywhere. Same shit with all these attempts to use ML to reconstruct or enhance data or interpolate missing data. You're ultimately choosing to train the AI to look at things a particular way and then getting results that follow the patterns it was trained on.

Meanwhile researchers are getting lazy - after all, why come up with some new optimized simulation or theoretical model when you can just have an AI bang its digital head against a wall until it gets something that looks like it could be vaguely right?

Anonymous

Anonymous Sat 02 Jan 2021 03:07:50 No.12534422 Report

Quoted By: >>12534462 >>12534726

>>12533287
You don't need to "know" how stuff works if you can predict how things behave this is literally how all science works. You could spend an eon to decode a machine learned algorithm but there wouldn't be much point in it. Inspecting the self driving cars algorithm for recognizing traffic lights would just inform you that the AI decides things by looking at the pixels which is how you also decide things.

Anonymous

Anonymous Sat 02 Jan 2021 03:11:03 No.12534428 Report

Quoted By:

Relevant blog post

https://reality.ai/ffts-and-stupid-deep-learning-tricks/

Anonymous

Anonymous Sat 02 Jan 2021 03:24:13 No.12534462 Report

Quoted By: >>12534503 >>12534757

>>12534422
If you are just frankensteining libraries together in order to create a functional software, you are not a programmer.

We don't have to recreate the universe for every pie we bake, but if everything comes in a microwaveable package, at one point we won't be able to competently recreate a pie when the company shuts down.

Anonymous

Anonymous Sat 02 Jan 2021 03:38:48 No.12534503 Report

Quoted By: >>12534726

>>12534462
>If you are just frankensteining libraries together in order to create a functional software, you are not a programmer.
Yes you are an machine learning expert worth about 10 programmers.

Anonymous

Anonymous Sat 02 Jan 2021 05:03:20 No.12534726 Report

Quoted By:

>>12534422
>>12534503
Both of these posts are business-oriented.

Anonymous

Anonymous Sat 02 Jan 2021 05:12:39 No.12534757 Report

Quoted By:

>>12534462
I've noticed this is the case for many ML type tasks. It's essentially just using numpy and scipy and hamfisting everything together.

Anonymous

Anonymous Sat 02 Jan 2021 09:29:08 No.12535535 Report

Quoted By:

>>12533287
The problem you described, unsolvable without ML, doesn't exist.

Please provide a problem which seems unsolvable.

Capcode	All Only User Posts Only Moderator Posts Only Admin Posts Only Developer Posts
Show Posts	All Only With Images Only Without Images
Deleted Posts	All Only Deleted Posts Only Non-Deleted Posts
Ghost Posts	All Only Ghost Posts Only Non-Ghost Posts
Post Type	All Only Sticky Threads Only Opening Posts Only Reply Posts
Results	All Grouped By Threads
Order	Latest Posts First Oldest Posts First

Your latest searches