【转载】关于CNN和RNN联系与区别的通俗解释

来自知乎网友的答案
作者：知乎用户
链接：https://www.zhihu.com/question/34681168/answer/84185831
来源：知乎
著作权归作者所有。商业转载请联系作者获得授权，非商业转载请注明出处。

Artificial neural networks use networks of activation units (hidden units) to map inputs to outputs. The concept of deep learning applied to this model allows the model to have multiple layers of hidden units where we feed output from the previous layers. However, dense connections between the layers is not efficient, so people developed models that perform better for specific tasks.

The whole “convolution” in convolutional neural networks is essentially based on the fact that we’re lazy and want to exploit spatial relationships in images. This is a huge deal because we can then group small patches of pixels and effectively “downsample” the image while training multiple instances of small detectors with those patches. Then a CNN just moves those filters around the entire image in a convolution. The outputs are then collected in a pooling layer. The pooling layer is again a down-sampling of the previous feature map. If we have activity on an output for filter a, we don’t necessarily care if the activity is for (x, y) or (x+1, y), (x, y+1) or (x+1, y+1), so long as we have activity. So we often just take the highest value of activity on a small grid in the feature map called max pooling.

If you think about it from an abstract perspective, the convolution part of a CNN is effectively doing a reasonable way of dimensionality reduction. After a while you can flatten the image and process it through layers in a dense network. Remember to use dropout layers! (because our guys wrote that paper :P)

I won’t talk about RNN for now because I don’t have enough experience working with them and according to my colleague nobody really knows what’s going on inside an LSTM…but they’re insanely popular for data with time dependencies.

Let’s talk RNN. Recurrent networks are basically neural networks that evolve through time. Rather than exploiting spatial locality, they exploit sequential, or temporal locality. Each iteration in an RNN takes an input and it’s previous hidden state, and produces some new hidden state. The weights are shared in each level, but we can unroll an RNN through time and get your everyday neural net. Theoretically RNN has the capacity to store information from as long ago as possible, but historically people always had problems with the gradients vanishing as we go back further in time, meaning that the model can’t be differentiated numerically and thus cannot be trained with backprop. This was later solved in the proposal of the LSTM architecture and subsequent work, and now we train RNNs with BPTT (backpropagation through time). Here’s a link that explains LSTMs really well: http://colah.github.io/posts/2015-08-Understanding-LSTMs/

Since then RNN has been applied in many areas of AI, and many are now designing RNN with the ability to extract specific information (read: features) from its training examples with attention-based models.