What is the purpose of batching in ML/DL?

[peak cedar]

Hello, I know sometimes, we split out dataset into batches, can someone explain why pls. I understood that maybe sometimes, we don't have enough memory and that's where we split into batches but apart that, are there other benefits pls

[torpid tangle]

im still a beginner but from what I know it would take a really long time if we fed it all the data at once and it stops the model from memorizing the patterns in noise but i think its mostly about the speed and efficiency im pretty sure

[river patio]

you can't feed all the data at once

training the data requires first load that onto ram/vram

generally speaking, if you have small dataset ( just for example ) you can feed all that once but again that goes into ram/vram

by using batches we can use all the size of ram/vram

[peak cedar]

yep I see, question though

how when we use batches, we prevent the model from memorizing patterns?

[river patio]

thats whole different question

learn about Neural network firsts

[peak cedar]

yep, I do have an overview of what they do

[peak cedar]

is it that it prevent overfitting but I don't understand how :c

[earnest yacht]

it pushes the model to focus on signal (consistent patterns across batches) rather than noise (idiosyncrasies of individual samples)

in full-batch: models can precisely fit all examples (including noisy ones) because gradients (in gradient descent) are pointing exactly toward minimizing total loss

in smaller batches: each gradient updates based on only a subset of data; the model learns more general patterns that are robust across these different smaller batches

I think this is what was meant by 'stops the model from memorizing the patterns in noise'