سلام
دوستانی که با معماری overfeat آشنایی دارند یا یک معماری دیگری هم اگر باشد که مانند این overfeat برای computation redundant یک راه حلی بیان کرده لطفا در مورد نحوه غلبه بر این توضیح دهند
One interesting contribution in Overfeat is multi-scale classification, which takes advantage of the spatial feature of convolutional neural nets to reduce computation.
So we can apply non-overlap convolutional and pooling operation on the inital feature map or image to get the results for its crops. The bonus here is that the redundant computation we talked about before is avoided. The comprise on the architecture here is non-overlap, no-overlap filters and non-overlap poolings.
If the redundant computation is avoided, we can crops with different size to train different classifiers in order to ensemble the results. This is called multi-scale classification in the paper.
یعنی چجوری با multi-scale classification به computation reduce رسید ؟ دوستانی که درک خوبی از این دارند لطفا توضیح دهند . یعنی اینکه non-overlapping خوبه ؟
https://arxiv.org/abs/1312.6229