Python Artificial Intelligence Projects for Beginners

[ 80 ]
Feed-forward neural networks
Most of the neural networks that we design are feed forward and fully connected. This
means that every neuron connects to every neuron in the next layer. The first layer receives
inputs and the last layer gives outputs. The structure of the network, meaning the neuron
counts and their connections, is decided ahead of time and cannot change, at least not
during training. Also, every input must have the same number of values. This means that
images, for example, may need to be resized to match the number of input neurons. The
number of neurons in each layer is that layer's shape:
(GGFHQTYCTFPGWTCNFGUKIP
Each individual neuron adds up the values it receives from the prior layer. Each connection
from one neuron to the next has a weight. When adding the inputs, the inputs are
multiplied by the respective weights. Each neuron also has an extra input called a 
bias
,
which is not connected to any other neurons. Once the weighted inputs have been added,
an activation function is applied to the sum.

---

Neural Networks
Chapter 4
[ 81 ]
There are several common activation functions, for example, the hyperbolic tangent, whose
shape is shown here:
*[RGTDQNKEVCPIGPV
The output of each neuron is whatever comes out of the activation function.
The connection waits in a network start random and are adjusted during training. The
purpose of training is to examine hundreds, or thousands, or even more example cases and
adjust the network's weights until the network is sufficiently accurate.
After training, we have a network structure that we have already defined, and all the
weights that were learned during training. As such, the following is true:
A trained neural network = Structure + Learned weights

---

Neural Networks
Chapter 4
[ 82 ]
This is shown here:
0GVYQTMUVTWEVWTGCHVGTVTCKPKPIYKVJYGKIJVU
Now the network is ready to use on new data outside the training set.

---

Neural Networks
Chapter 4