Welcome back, in the last video you learned about the depth-wise separable convolution. Let's now put this into a neural network in order to build the MobileNet. The idea of MobileNet is everywhere that you previously have used an expensive convolutional operation. You can now instead use a much less expensive depthwise separable convolutional operation, comprising the depthwise convolution operation and the pointwise convolution operation. The MobileNet v1 paper had a specific architecture in which it use a block like this, 13 times. It would use a depthwise convolutional operation to genuine outputs and then have a stack of 13 of these layers in order to go from the original raw input image to finally making a classification prediction. Just to provide a few more details after these 13 layers, the neural networks last few layers are the usual Pooling layer, followed by a fully connected layer, followed by a Softmax in order for it to make a classification prediction. This turns out to perform well while being much less computationally expensive than earlier algorithms that used a normal convolutional operation. In this video, I want to share with you one more improvements on this basic MobileNet architecture, which is the MobileNets v2 architecture. I'm going to go through this quite quickly to just give you a rough sense of what the algorithm does without digging into all of the details. But this is a paper published by Mark Sandler and his colleagues that make mobile network even better. In MobileNet v2, there are two main changes. One is the addition of a residual connection. This is just a residual connections that you learned about in the ResNet videos. This residual connection or skip connection, takes the input from the previous layer and sums it or passes it directly to the next layer, does allow ingredients to propagate backward more efficiently. The second change is that it also as an expansion layer, which you learn more about on the next slide, before the depthwise convolution, followed by the pointwise convolution, which we're going to call projection in a point-wise convolution. Is really the same operation, but we'll give it a different name, for reasons that you see on the next slide. What MobileNet v2 does, is it uses this block and repeats this block some number of times, just like we had 13 times here, the MobileNet v2 architecture happened to choose to do this 17 times, surpass the inputs through 17 of these blocks, and then finally ends up with the usual pooling fully-connected softmax, in order to make a classification prediction. But the key idea is really how blocks like this, or like this reduces computational cost. This block over here is also called the bottleneck block. Let's dig into the details of how the MobileNet v2 block works. Given an input that is say, n by n by three, the MobileNet v2 bottleneck will pass that input via the residual connection directly to the output, just like in the Resnet. Then in this main non-residual connection parts of the block, you'll first apply an expansion operator, and what that means is you'll apply a 1 by 1 by n c. In this case, one by one by three-dimensional filter. But you apply a fairly large number of them, say 18 filters, so that you end up with an n by n by 18-dimensional block over here. A factor of expansion of six is quite typical in MobileNet v2 which is why your inputs goes from n by n by three to n by n by 18, and that's why we call it an expansion as well, it increases the dimension of this by a factor of six, the next step is then a depthwise separable convolution. With a little bit of padding, you can then go from n by n by 18 to the same dimension. In the last video, we went from 6 by 6 by 3 to 4 by 4 by 3 because we didn't use padding. But with padding, you can maintain the n by n by 18 dimension, so it doesn't shrink when you apply the depthwise convolution. Finally, you apply a pointwise convolution, which in this case means convolving with a 1 by 1 by 18-dimensional filter. If you have, say, three filters and c prime filters, then you end up with an output that is n by n by 3 because you have three such filters. In this last step, we went from n by n by 18 down to n by n by three, and in the MobileNet v2 bottleneck block. This last step is also called a projection step because you're projecting down from n by n by 18 down to n by n by 3. You might be wondering, why do we meet these bottleneck blocks? It turns out that the bottleneck block accomplishes two things, One, by using the expansion operation, it increases the size of the representation within the bottleneck block. This allows the neural network to learn a richer function. There's just more computation over here. But when deploying on a mobile device, on edge device, you will often be heavy memory constraints. The bottleneck block uses the pointwise convolution or the projection operation in order to project it back down to a smaller set of values, so that when you pass this the next block, the amount of memory needed to store these values is reduced back down. The clever idea, The cool thing about the bottleneck block is that it enables a richer set of computations, thus allow your neural network to learn richer and more complex functions, while also keeping the amounts of memory that is the size of the activations you need to pass from layer to layer, relatively small. That's why the MobileNet v2 can get a better performance than MobileNet v1, while still continuing to use only a modest amount of compute and memory resources. Putting It all together again, here's our earliest slide describing MobileNet v1 and v2. The MobileNet v2 architecture will repeat this bottleneck block 17 times. Then if your goal is to make a classification, then as usual will have this pooling fully-connected softmax layer to generate the classification output. You now know the main ideas of MobileNet v1 and v2. If you want to see a few additional details such as the exact dimensions of the different layers as you flow the data from left to right. You can also take a look at the paper by Mark Sandler and others as well. Congrats on making it to the end of this video, you now know how MobileNet v1 and v2 work by using depthwise separable convolutions and by using the bottleneck blocks that you saw in this video. If you are thinking of building efficient neural networks, there's one other idea to found very useful, and I hope you will learn about as well, which is efficient nets. Let's touch on that idea very briefly in the next video.
