Hello, there is something bugs me about TTA (Test Time Augmentation). Since we do not use validation loss for back propagation (as I understand) how come it improves our results? Is it stupid to ask? Thanks for any help. I might totally misunderstood the topic as well.
Hey @nikem test time augmentation is not a part of the training process so there is no backpropagation. We never let the model learn from the validation set, we only do inference on it. In TTA, we show the model different simple augementations of an image and take its predictions. These predictions are aggregated to make a better prediction.
In a more intuitive sense, we use the augmentations to maybe excite or activate some of the neurons that help make a more accurate prediction. For example, the fur of the bear might help us decide if the bear is real or if it’s just a teddy. This may only be clear when we zoom in on the test image.
It’s an ensemblig technique where instead of using multiple models, we use multiple different augmentations of the image.
Thanks dhruv. But that means we don’t really improve the model. Just the results,right?
@nikem yeah you can say that. We are making more use of our trained model!