I was watching a video which tells a bit about reinforcement learning, and I learnt that If the robot makes wrong movement then they train the network with negative learning rate. From this method, something came to my mind.
My question is "Can I use a wrong data to train a neural network?".
To illustrate the method, I'll be using the eye tracker project that I'm working on right now. In my project There are photos and the points that corresponds the locations that I m looking to at that photo. Its like grid (9, 16). If I look to the middle of the screen, it means the output is (4, 7.5). if I look left up side of the screen it means (0,0). Normally for a photo that I'm looking into the middle, we use that photo as input and (4, 7.5) as output to train network using positive learning rate. Now let me rephrase the question. Can I train a model giving a photo that I'm looking into the middle as input and (0,0) as output(label) using negative learning rate?
Thank you, If I made a mistake against the rules of stackoverflow, I'm so sorry. I'll be waiting your valuable answers.
Edit: this is a conversation between me and someone from stackoverflow, I'll let you read, hope you get a point.
-> Yes, you can. But, what would be the reason of passing a wrong ground truth to your training process? – Neb 14 hours ago
-> If I have no various data to train, I can create more data via this method to increase the certainty when I use squared error loss. But I have doubts about this method. for example lets assume we have a photo named 'X' and its label is (5,5). at first epoch, Let the model gives (2,2) for photo 'X'. if I try to train network with a photo X and label -> (4,4) using negative learning rate, it might send away the point from (2,2) to (1,1) whereas we expect it to send the point (2,2) to (5,5). Did you get what I meant? – Faruk Nane 14 hours ago
-> You are right. Using a negative learning rate and a wrong ground truth will not necessarly make the learning process converge to the optimal value for your net's parameters – Neb 13 hours ago
-> So can I say that "when I'm sure that the absolute error for each case is less than 2, I can use this method using points away 2 units." So It'll make the outputs closer to the target point. I don't really know if we can easily say that. because we consider this method as if there are only 2 parameters which is the output point. However a model has many parameters so It might affect so differently. My brain is so confused. I think this might be an academic work, right? – Faruk Nane 13 hours ago
-> Well, it is difficult to suggests you the path to follow without knowing the exact specifics of your problem. In any case, if you're trying to solve this problem for fun or self-improvement, I'd suggest you to experiment with the solutions you came up with and see if they works. – Neb 13 hours ago
//EDIT: UP UP