Neural Networks in a nutshell . . .

Feed Forward calculations

Input values enter the model on the far left. The input values are then "plugged-in" to a formula (i.e., Output Formula on page 254) and the resulting output value is passed to the next node to the right. This process continues until the output node (or nodes) is reached on the far right of the network.

Back Propagation calculations

The output from the network is then compared to the Training Data and an Error Value (or difference) is computed using the Error formula (bottom of page 257) . The error from the Error Formula is then used to determine new weights and new bias values (formulas also on bottom of page 257). The new weights and new bias values are then "plugged-in" to the Output Formulas in the network. After the Output Formulas have been adjusted with the new weights and bias values, the Feed Forward process is then repeated.

Multiple Passes Made through Network

The whole process above continues to repeat until the weights and bias values no longer change (i.e., lowest error reached) or a specified number of passes (i.e., epochs) through the network has been reached. In summary, Neural Networks "learn from their mistakes" until there is no more learning that can be done (which usually results in "overlearning" or what we call "overfitting), or until a set number of passes through the network (i.e., epochs) is reached.

Focus on the "big picture" process of how Neural Networks "learn" and try not to get overwhelmed by all of the specific formulas involved.