Hey Jessy Scaria,
This question is like asking how many particles of powdered sugar do you need in your coffee/tea to get a decent taste. Because 1000 particles is also going to taste the same to you as the 10010 particles of sugar.
Because there is no correct one digit answer to this question. You need a range of number. and that number depends the size of your neural net and the total number of categories you want to classify.
Look at this problem from a different angle. Initially, never think how much dataset do you need.
- Start by thinking how many classifications do you want your neural network to do. (aka Diversity of your Neural Network)
- I know it doesn’t sound right but don’t worry about the resolution of the pictures unless you are making a product (because generally all the images that people download from google have decent resolution)
- If your demand of categories (point1) is low. You don’t need a big neural net. And small neural net don’t need alot of data to train on anyways.
- Also on a side note, you can just increase the number of epochs if you have less data. Just don’t increase it too much, eventually it will over-fit your network.
- The most important point, even if you have small dataset (Just train the network anyways) then use trial and error method to get a good enough trained NN.
I think the ans is satisfactory enough, If not — Ask again.
If you are training — try these tutorials:
Disclaimer: Read further at your own risk.
Now let try to make this comment more interesting. The question you have asked is one of the most asked question in Machine Learning field.
The right question is not “How many Images do I need to train a neural network?”
The question is “how many Images can a specific neural network handle. And when will it the neural network will reach it limit where its size doesn’t allow it to learn more, pushing the user to increase its size.
and when Increasing the size — should we increase the no. of nodes in the already existing layers or should we increase the layers. And what the difference between both the process.
BBye
