Want to know more about Robots BLOG POST: vaishviksatyam.wordpress.com Host and Creator - Vaishvik Satyam WEBSITE: vaishviksatyam.wordpress.com QUESTIONS: themechatronicsolutions@gmail.com
You lost me at "If I give lesser epochs, the line will split the curve"... :P I think what he tries to say is, one epoch = one forward pass and one backward pass of ALL the training examples. Now that we know what an epoch is, then how many number of epochs do you want to run thorugh the neural network so it can train its weight properly? ANSWER: f you run ALL the training data several times, you might end up with overtrained weights. You want to find a balance between "less epochs" and "alot of epochs". Example, if "less epochs" is five runs (epochs), and "alot of epochs" is 100 runs. Then maybe the perfect balance to categorize new unseen images will be 40 runs.More than 40 runs (epochs) would have given you overfitting, and less epochs would have fiven you underfitting. How do you find the optimal number of epochs? Answer: I think it is just trial and error. Trying different number of epochs, and then evaluate how well it did on the new unseen data.
Instead of just trial and error, it is possible to plot the validation loss & training loss. Training loss will almost always go down with more epochs. We have to keep a lookout for validation loss. As long as it is decreasing with epochs, we are good to train.
An epoch is normally a measure of time, in deep learning, it's a measure of the number of times the entire dataset is passed through the model structure. You can (will) overfit the data if you loop through the dataset too many times. The goal is to stop training when your model has learned to generalize without memorizing. How will you recognize that point? Generally, you will graph the loss from the training data and test / unseen data. While a model is "learning" these two curves will be roughly parallel. When the training loss curve continues to decline but the test / unseen data loss curve is no longer roughly parallel, you have reached the point of overtraining. I hope this makes sense.
We can divide the dataset of 2000 examples into batches of 500 then it will take 4 iterations to complete 1 epoch. ( A bit of thinking and you'll understand everything)
You are judging too quickly. :-) they were not well so I got them to my place. Since they got better I set them free again. Good news is they sometimes come back, chirp a little and go back.
Epoch is a bunch of data given to the neutral network from whole lot of data.. example.. while eating one full plate of rice, each spoon you eat is an epoch. You can only take little amount. Else you will vomit. If you take too small it will take a while to complete and you will be hungry for long.