 So we've taken all these individual Lego blocks now, we've added them into structure, but you can think of them as just free-floating in space. They're not connected yet. It doesn't know which comes first and which output to connect to which input of which other block. To do that, we connect them. So we go to our classifier. We can go through and individually connect the output of each to the input of the next all the way through. What is convenient to do in neural networks, often there's a long train of elements where the output of one is connected to the input of the next. And so we have a convenience function here, connect sequence, where we pass it a list of names of blocks to connect in a sequence. So we connect our training data to our convolution, to our bias, to our activation function, to the next convolution, the next bias, the next activation function, to our pooling, our flatten, our linear and our bias and our logistic function, and then to the prediction and then to the loss function. You can see in the structure diagram how there is a line starting at the training data tracing all the way through to the loss function. There are a few connections now that haven't yet been accounted for. So on these, we have to go through and make these individually. By default, the connect sequence function assumes that the zeroth port on the tail of that connection and the zeroth port on the block at the head of that connection are what's being connected. Each block can have as many connection ports as we want. And we just have to specify when we're connecting what the index of the port is that we're connecting. So for instance, if we're connecting our training data to our one hot block, our zeroth port on the training data is already occupied. It's already connected to the convolution block. So we want to use port one on the training data block. So we specify I port tail. So on the connection, this is the tail of the connection. The head of the connection is going to our one hot block. It'll still go to the zeroth port so we can let it default to zero. Now when we want to connect our one hot block to our loss function, the input port on the loss function, the zeroth port there is already occupied. So we want to specify I port head equals one. The head of that connection is on the loss block and we want it to be port one. Similarly, when we're connecting our prediction block, so our copy of our logistic results to our hardmax, our zeroth port on the prediction block is already occupied. So on the tail of that connection, we want to make sure and use port one. So we specify I port tail equals one. With these three remaining connections, we've completed our structure diagram as we see it here. So now we have built our graph. We've built our model.